Upload full repo excluding dump_40, dump_100, precomputed_tokens, precomputed_data

.gitattributes

@@ -33,3 +33,18 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+Data/OOD_texts.txt filter=lfs diff=lfs merge=lfs -text
+Data/finalfinal.txt filter=lfs diff=lfs merge=lfs -text
+Data/harry.txt filter=lfs diff=lfs merge=lfs -text
+Data/normalize_log.csv filter=lfs diff=lfs merge=lfs -text
+Data/syspin.txt filter=lfs diff=lfs merge=lfs -text
+Data/texts/fin.txt filter=lfs diff=lfs merge=lfs -text
+Data/train_3.txt filter=lfs diff=lfs merge=lfs -text
+Data/train_3_filtered.txt filter=lfs diff=lfs merge=lfs -text
+Data/train_3_filtered_1.txt filter=lfs diff=lfs merge=lfs -text
+Data/train_3_filtered_dur.txt filter=lfs diff=lfs merge=lfs -text
+Data/train_3_filtered_dur_fixed.txt filter=lfs diff=lfs merge=lfs -text
+Data/train_3_filtered_dur_fixed_copy.txt filter=lfs diff=lfs merge=lfs -text
+Data/train_3_filtered_dur_fixed_speakers.txt filter=lfs diff=lfs merge=lfs -text
+infer_outputs/sample_01.wav filter=lfs diff=lfs merge=lfs -text
+infer_outputs/story.wav filter=lfs diff=lfs merge=lfs -text

Configs/config.yml

@@ -0,0 +1,114 @@
+log_dir: "again_wavlm"
+first_stage_path: ""
+save_freq: 2
+log_interval: 10
+device: "cuda"
+epochs_1st: 150 # number of epochs for first stage training (pre-training)
+epochs_2nd: 100 # number of peochs for second stage training (joint training)
+batch_size: 32
+max_len: 8000 # maximum number of frameSs
+pretrained_model: "again_wavlm/epoch_flow_00108.pth"
+second_stage_load_pretrained: false # set to true if the pre-trained model is for 2nd stage
+load_only_params: false # set to true if do not want to load epoch numbers and optimizer parameters
+
+F0_path: "Utils/JDC/epoch_000401_final.pth"
+ASR_config: "Utils/ASR/config.yml"
+ASR_path: "Utils/ASR/epoch_000801.pth"
+PLBERT_dir: 'Utils/PLBERT/'
+
+data_params:
+  
+  #train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  OOD_data: "Data/OOD_texts.txt"
+  min_length: 50 # sample until texts with this size are obtained for OOD texts
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 4096
+    win_length: 2400
+    hop_length: 600
+
+model_params:
+  multispeaker: false
+
+  dim_in: 64 
+  hidden_dim: 512
+  max_conv_dim: 512
+  n_layer: 3
+  n_mels: 128
+
+  n_token: 178 # number of phoneme tokens
+  max_dur: 75 # maximum duration of a single phoneme
+  style_dim: 64 # style vector size
+  style_dim_ac: 128 # style vector size
+  dropout: 0.1
+
+  # config for decoder
+  decoder: 
+      type: 'ringformer' # either hifigan or istftnet
+      resblock_kernel_sizes: [3,7,11]
+      upsample_rates :  [12,10]
+      upsample_initial_channel: 512
+      resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]]
+      upsample_kernel_sizes: [24, 20]
+      gen_istft_n_fft: 30
+      gen_istft_hop_size: 5
+      
+  # speech language model config
+  slm:
+      model: 'microsoft/wavlm-base-plus'
+      sr: 16000 # sampling rate of SLM
+      hidden: 768 # hidden size of SLM
+      nlayers: 13 # number of layers of SLM
+      initial_channel: 64 # initial channels of SLM discriminator head
+  
+  # style diffusion model config
+  diffusion:
+    embedding_mask_proba: 0.2
+    # transformer config
+    transformer:
+      num_layers: 4
+      num_heads: 4
+      head_features: 32
+      multiplier: 4
+
+    # diffusion distribution config
+    dist:
+      sigma_data: 0.5 # placeholder for estimate_sigma_data set to false
+      mean_data: 0.1
+      acoustic: 0.46
+      prosodic: 0.46
+      pitch: 0.46
+      
+      acoustic_mean: 0.4
+      prosodic_mean: 0.4
+      pitch_mean: 0.4
+      estimate_sigma_data: true # estimate sigma_data from the current batch if set to true
+      mean: -0.0
+      std: 0.3
+  
+loss_params:
+
+    lambda_mel: 10. # mel reconstruction loss
+    lambda_gen: 1. # generator loss
+    lambda_slm: 1. # slm feature matching loss
+    TMA_epoch: 10 # TMA starting epoch (1st stage)
+    lambda_mag: 0.5
+optimizer_params:
+  lr: 0.0001 # general learning rate
+  bert_lr: 0.00001 # learning rate for PLBERT
+  ft_lr: 0.00002 # learning rate for acoustic modules
+  
+slmadv_params:
+  min_len: 400 # minimum length of samples
+  max_len: 600 # maximum length of samples
+  batch_percentage: 0.5 # to prevent out of memory, only use half of the original batch size
+  iter: 15 # update the discriminator every this iterations of generator update
+  thresh: 5 # gradient norm above which the gradient is scaled
+  scale: 0.01 # gradient scaling factor for predictors from SLM discriminators
+  sig: 1.5 # sigma for differentiable duration modeling
+  

Configs/config_codec.yml

@@ -0,0 +1,133 @@
+# Codec Training Configuration
+# ==============================================================================
+# This config is for training the RVQ-VAE codec vocoder
+# Based on config.yml but with codec-specific settings
+# ==============================================================================
+
+log_dir: "codec_training"
+save_freq: 2
+log_interval: 10
+device: "cuda"
+
+# Training epochs
+epochs_codec: 100
+
+batch_size: 5  # Reduced due to additional codec memory
+max_len: 8000
+
+# Pretrained models (optional - for warm starting)
+pretrained_model: "codec_training/epoch_codec_00018.pth"  # Base model for text/style encoders
+pretrained_codec: "codec_training/epoch_codec_00018.pth"  # Set to load a pretrained codec
+
+# Quantizer warmup (gradually increase number of quantizers)
+quantizer_warmup_steps: 10000
+min_quantizers: 1
+
+# Data
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  OOD_data: "Data/OOD_texts.txt"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 4096
+    win_length: 2400
+    hop_length: 600
+
+# Model parameters
+model_params:
+  multispeaker: false
+
+  dim_in: 64 
+  hidden_dim: 512
+  max_conv_dim: 512
+  n_layer: 3
+  n_mels: 128
+
+  n_token: 178
+  max_dur: 75
+  style_dim: 64
+  style_dim_ac: 128
+  dropout: 0.1
+
+  # === CODEC SETTINGS ===
+  # Latent dimension for RVQ
+  codec_latent_dim: 512
+  
+  # Number of RVQ codebooks (more = higher quality, more tokens)
+  num_quantizers: 8
+  
+  # Codebook size (vocabulary size per quantizer)
+  codebook_size: 4096
+  
+  # Compression strides (total compression = product)
+  # [2, 2, 2] = 8x compression
+  codec_strides: [1]
+
+  # Decoder/Vocoder config
+  decoder: 
+    type: 'ringformer'
+    resblock_kernel_sizes: [3, 7, 11]
+    upsample_rates: [12, 10]
+    upsample_initial_channel: 512
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    upsample_kernel_sizes: [24, 20]
+    gen_istft_n_fft: 30
+    gen_istft_hop_size: 5
+      
+  # SLM for perceptual loss
+  slm:
+    model: 'microsoft/wavlm-base-plus'
+    sr: 16000
+    hidden: 768
+    nlayers: 13
+    initial_channel: 64
+  
+  # Style diffusion (not used in codec training but needed for model building)
+  diffusion:
+    embedding_mask_proba: 0.2
+    transformer:
+      num_layers: 4
+      num_heads: 4
+      head_features: 32
+      multiplier: 4
+    dist:
+      sigma_data: 0.5
+      mean_data: 0.1
+      acoustic: 0.46
+      prosodic: 0.46
+      pitch: 0.46
+      acoustic_mean: 0.4
+      prosodic_mean: 0.4
+      pitch_mean: 0.4
+      estimate_sigma_data: true
+      mean: -0.0
+      std: 0.3
+
+# Optimizer
+optimizer_params:
+  lr: 1.0e-4
+  pct_start: 0.0
+
+# Loss weights
+loss_params:
+  # Waveform losses
+  lambda_mel: 10.0        # Multi-resolution STFT
+  lambda_gen: 1.0         # GAN generator loss
+  lambda_slm: 1.0         # Speech language model loss
+  lambda_mag: 0.5         # Magnitude/phase loss
+  
+  # Codec-specific losses
+  lambda_codec: 1.0       # Overall codec loss weight
+  lambda_pitch: 1.0       # Pitch reconstruction
+  lambda_energy: 1.0      # Energy reconstruction  
+  lambda_text: 2.0        # Text embedding reconstruction
+  lambda_commitment: 0.25 # VQ commitment loss
+  lambda_diversity: 0.02  # Codebook diversity regularization
+  
+  # GAN training starts after this epoch
+  TMA_epoch: 15

Configs/config_codec_hybrid.yml

@@ -0,0 +1,118 @@
+# Hybrid Codec Training Configuration
+# ==============================================================================
+# HYBRID APPROACH:
+#   - Text stays CONTINUOUS (not quantized) for better intelligibility
+#   - Pitch, Energy, Style are QUANTIZED via FSQ for prosodic compression
+#   - Everything is downsampled to compressed representation
+# ==============================================================================
+
+log_dir: "hybrid_codec_training"
+save_freq: 1
+log_interval: 10
+device: "cuda"
+
+# Training epochs
+epochs_codec: 100
+
+batch_size: 4
+max_len: 8000
+
+# Pretrained models
+pretrained_model: "hybrid_codec_training/epoch_hybrid_codec_00017.pth"  # Base model for text/style encoders
+pretrained_hybrid_codec: "hybrid_codec_training/epoch_hybrid_codec_00017.pth"  # Set to load a pretrained hybrid codec
+
+# Data
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  OOD_data: "Data/OOD_texts.txt"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 4096
+    win_length: 2400
+    hop_length: 600
+
+# Model parameters
+model_params:
+  multispeaker: false
+
+  dim_in: 64 
+  hidden_dim: 512
+  max_conv_dim: 512
+  n_layer: 3
+  n_mels: 128
+
+  n_token: 178
+  max_dur: 75
+  style_dim: 64
+  style_dim_ac: 128
+  dropout: 0.1
+
+  # === HYBRID CODEC SETTINGS ===
+  # Prosody latent dimension (for pitch/energy/style only)
+  # Text goes through separately and stays continuous
+  prosody_latent_dim: 256
+  
+  # Hidden dimension for decoder
+  codec_hidden_dim: 512
+  
+  # FSQ levels for prosody quantization
+  # Product = codebook size (4^6 = 4096 codes)
+  # Only prosody is quantized - text stays continuous!
+  fsq_levels: [4, 4, 4, 4, 4]
+  
+  # Compression strides (applies to BOTH prosody and text)
+  # [2, 2] = 4x compression
+  # Text is downsampled but NOT quantized
+  codec_strides: [2]
+
+  # Decoder/Vocoder config
+  decoder: 
+    type: 'ringformer'
+    resblock_kernel_sizes: [3, 7, 11]
+    upsample_rates: [12, 10]  # Product: 480, matches 600 * 4 / 5
+    upsample_initial_channel: 512
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    upsample_kernel_sizes: [24,20]
+    gen_istft_n_fft: 30
+    gen_istft_hop_size: 5
+      
+  # SLM for perceptual loss
+  slm:
+    model: 'microsoft/wavlm-base-plus'
+    sr: 16000
+    hidden: 768
+    nlayers: 13
+    initial_channel: 64
+  
+  # Style diffusion (not used in codec training but needed for model building)
+  diffusion:
+    embedding_mask_proba: 0.2
+    transformer:
+      num_layers: 4
+      num_heads: 4
+      head_features: 32
+      multiplier: 4
+    dist:
+      sigma_data: 0.5
+      mean_data: 0.1
+
+# Loss weights
+loss_params:
+  TMA_epoch: 10
+  lambda_mel: 20.0
+  lambda_gen: 1.0
+  lambda_slm: 1.0
+  lambda_mag: 0.2
+  lambda_entropy: 50.0  # Increased to encourage codebook utilization
+  lambda_f0: 1.0
+  lambda_style: 1.0
+
+# Optimizer
+optimizer_params:
+  lr: 0.0001
+  pct_start: 0.0

Configs/config_codec_hybrid_temporal.yml

@@ -0,0 +1,110 @@
+# Hybrid Codec with WINDOWED Temporal Style Training Configuration
+# ==============================================================================
+# KEY DIFFERENCE:
+#   - Style encoder uses large overlapping windows (~3sec) instead of T//16
+#   - Captures timbre/speaker characteristics, NOT phoneme-level detail
+#   - Output: [B, num_windows, 32] where num_windows depends on audio length
+#   - Prevents style encoder from overfitting to phoneme content
+# ==============================================================================
+
+log_dir: "hybrid_temporal_codec_training"
+save_freq: 1
+log_interval: 10
+device: "cuda"
+
+# Training epochs
+# Pre-TMA: 6 epochs (warmup), TMA: 14 epochs (decay)
+epochs_codec: 20
+
+batch_size: 6
+max_len: 7000
+
+# n_mels - 40 by default for temporal style encoder
+n_mels: 40
+
+# Pretrained models
+pretrained_model: ""
+
+# Data
+data_params:
+  train_data: "../my_tts_dataset_v2/train_list_filtered_german.txt"
+  val_data: "../my_tts_dataset_v2/val_list_filtered_german.txt"
+  root_path: "../my_tts_dataset_v2"
+  OOD_data: "Data/OOD_texts.txt"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 2048
+    win_length: 1764
+    hop_length: 441
+
+# Model parameters
+model_params:
+  multispeaker: false
+
+  dim_in: 64 
+  hidden_dim: 512
+  max_conv_dim: 512
+  n_layer: 3
+  n_mels: 40  # 40 mels for temporal style encoder
+
+  n_token: 176
+  max_dur: 75
+  style_dim: 32  # Windowed timbre style dimension (spatial)
+  style_dim_ac: 128  # Pooled style dimension for comparison
+  dropout: 0.1
+
+  # === HYBRID TEMPORAL CODEC SETTINGS ===
+  # Prosody latent dimension
+  prosody_latent_dim: 512
+  
+  # Hidden dimension for decoder
+  codec_hidden_dim: 512
+  
+  # FSQ levels for prosody quantization
+  fsq_levels: [4, 4, 4, 4, 4,4]
+  #fsq_levels: [8, 8, 8, 8]
+  # Compression strides
+  codec_strides: [2]
+
+  # Decoder/Vocoder config
+  decoder: 
+    type: 'ringformer'
+    resblock_kernel_sizes: [3, 7, 11]
+    upsample_rates: [7, 7]
+    upsample_initial_channel: 512
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    upsample_kernel_sizes: [14,14]
+    gen_istft_n_fft: 36
+    gen_istft_hop_size: 9
+      
+  # SLM for perceptual loss
+  slm:
+    model: 'microsoft/wavlm-base-plus'
+    sr: 16000
+    hidden: 768
+    nlayers: 13
+    initial_channel: 64
+  
+
+
+# Loss weights
+loss_params:
+  TMA_epoch: 4  # TMA starts at epoch 4 (after 4 warmup epochs)
+  lambda_mel: 10.0
+  lambda_gen: 1.0
+  lambda_slm: 1.0
+  lambda_mag: 0.2
+  lambda_entropy: 50.0
+  lambda_f0: 1.0
+  lambda_style: 2.0
+
+# Optimizer with LR scheduling
+# Schedule: decay from 1e-4 to 1e-5 during pre-TMA (epochs 0-6), then constant 1e-5 during TMA (epochs 6-18)
+optimizer_params:
+  lr: 1e-5           # Starting learning rate (pre-TMA)
+  min_lr: 1e-5      # Final LR (TMA phase, constant)
+  decay_epochs: 6      # Decay happens during pre-TMA phase
+  pct_start: 0.33      # 6/18 = 0.33 of training is decay phase

Configs/config_codec_mel_speaker.yml

@@ -0,0 +1,112 @@
+# Mel Speaker Codec Training Configuration
+# ==============================================================================
+# KEY FEATURES:
+#   - Input: mel spectrogram + pitch (NO text embeddings, NO energy)
+#   - Learnable speaker embeddings: nn.Embedding(num_speakers, speaker_dim)
+#   - Speaker conditioning via AdaIN1d throughout decoder
+#   - F0 prediction from latent representation
+#   - codec_strides=[4] for ~25Hz token rate
+#
+# This codec encodes mel spectrogram and pitch into a latent representation,
+# quantizes it with FSQ, predicts pitch, and decodes to waveform using
+# speaker embeddings for voice identity.
+# ==============================================================================
+
+log_dir: "mel_speaker_codec_training"
+save_freq: 1
+log_interval: 10
+device: "cuda"
+
+# Training epochs
+epochs_codec: 50
+
+batch_size: 8
+max_len: 6000
+
+# n_mels - for mel spectrogram input
+n_mels: 100
+
+# Dump directory for preprocessed data
+dump_dir: "./dump_100"
+
+# Pretrained models
+pretrained_model: "mel_speaker_codec_training/epoch_mel_speaker_codec_00002.pth"
+load_only_params: false
+
+# Data
+data_params:
+  train_data: "../my_tts_dataset_v2/train_list_filtered_german.txt"
+  val_data: "../my_tts_dataset_v2/val_list_filtered_german.txt"
+  root_path: "../my_tts_dataset_v2"
+  OOD_data: "Data/OOD_texts.txt"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 2048
+    win_length: 1764
+    hop_length: 441
+
+# Model parameters
+model_params:
+  # === SPEAKER EMBEDDING SETTINGS ===
+  num_speakers: 11      # Number of speakers (IDs 0-10)
+  speaker_dim: 64      # Speaker embedding dimension
+
+  # === MEL INPUT SETTINGS ===
+  n_mels: 100            # Mel spectrogram bins
+
+  # === LATENT/CODEC SETTINGS ===
+  latent_dim: 512       # Latent dimension
+  hidden_dim: 512       # Hidden dimension for encoder/decoder
+  
+  # FSQ levels for quantization
+  fsq_levels: [4, 4, 4, 4, 4, 4]  # 4096 codes
+  #fsq_levels: [4, 4, 4, 8]
+  # Compression strides - 4x for ~25Hz token rate
+  # 44100 / 441 / 4 = 25 Hz
+  codec_strides: [2]
+
+  # === PHONEME PREDICTOR SETTINGS ===
+  n_phonemes: 176       # Phoneme vocabulary size
+  num_languages: 10     # Number of languages for conditioning
+  language_dim: 64      # Language embedding dimension
+
+  # Decoder/Vocoder config (matches config_codec_speaker.yml)
+  decoder: 
+    type: 'ringformer'
+    resblock_kernel_sizes: [3, 7, 11]
+    upsample_rates: [7, 9]
+    upsample_initial_channel: 512
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    upsample_kernel_sizes: [14, 18]
+    gen_istft_n_fft: 28
+    gen_istft_hop_size: 7
+      
+  # SLM for perceptual loss
+  slm:
+    model: 'microsoft/wavlm-base-plus'
+    sr: 16000
+    hidden: 768
+    nlayers: 13
+    initial_channel: 64
+
+
+# Loss weights
+loss_params:
+  TMA_epoch: -1         # TMA (adversarial) starts at epoch 4
+  lambda_mel: 20.0      # STFT reconstruction loss
+  lambda_gen: 1.0       # Generator adversarial loss
+  lambda_slm: 1.0       # WavLM perceptual loss
+  lambda_mag: 0.2       # Magnitude/phase loss
+  lambda_entropy: 50.0  # FSQ entropy/commitment loss
+  lambda_f0: 1.0        # F0 prediction loss
+  lambda_phoneme: 0.5   # Phoneme prediction loss (frame-level CE)
+
+# Optimizer
+optimizer_params:
+  lr: 1e-5
+  min_lr: 1e-5
+  decay_epochs: 15
+  pct_start: 0.33

Configs/config_codec_speaker.yml

@@ -0,0 +1,110 @@
+# Speaker Codec Training Configuration
+# ==============================================================================
+# KEY FEATURES:
+#   - Learnable speaker embeddings: nn.Embedding(11, 128)
+#   - Speaker IDs 0-10 for 11 speakers
+#   - Speaker conditioning via AdaIN1d throughout decoder
+#   - NO mel-based style encoder - purely speaker ID based
+# ==============================================================================
+
+log_dir: "speaker_codec_training"
+save_freq: 1
+log_interval: 10
+device: "cuda"
+
+# Training epochs
+epochs_codec: 20
+
+batch_size: 6
+max_len: 7000
+
+# n_mels - for mel spectrogram extraction (used for loss calculation)
+n_mels: 40
+
+# Pretrained models
+pretrained_model: ""
+load_only_params: false
+
+# Data
+data_params:
+  train_data: "../my_tts_dataset_v2/train_list_filtered_german.txt"
+  val_data: "../my_tts_dataset_v2/val_list_filtered_german.txt"
+  root_path: "../my_tts_dataset_v2"
+  OOD_data: "Data/OOD_texts.txt"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 2048
+    win_length: 1764
+    hop_length: 441
+
+# Model parameters
+model_params:
+  multispeaker: true  # Enable multi-speaker mode
+
+  dim_in: 64 
+  hidden_dim: 512
+  max_conv_dim: 512
+  n_layer: 3
+  n_mels: 40
+
+  n_token: 176
+  max_dur: 75
+  dropout: 0.1
+
+  # === SPEAKER EMBEDDING SETTINGS ===
+  num_speakers: 11      # 11 speakers: IDs 0-10
+  speaker_dim: 128      # 128-dim learnable embeddings
+
+  # === CODEC SETTINGS ===
+  # Prosody latent dimension
+  prosody_latent_dim: 512
+  
+  # Hidden dimension for decoder
+  codec_hidden_dim: 512
+  
+  # FSQ levels for prosody quantization
+  fsq_levels: [4, 4, 4, 4, 4, 4]
+  
+  # Compression strides
+  codec_strides: [4]
+
+  # Decoder/Vocoder config
+  decoder: 
+    type: 'ringformer'
+    resblock_kernel_sizes: [3, 7, 11]
+    upsample_rates: [7, 9]
+    upsample_initial_channel: 512
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    upsample_kernel_sizes: [14, 18]
+    gen_istft_n_fft: 28
+    gen_istft_hop_size: 7
+      
+  # SLM for perceptual loss
+  slm:
+    model: 'microsoft/wavlm-base-plus'
+    sr: 16000
+    hidden: 768
+    nlayers: 13
+    initial_channel: 64
+
+
+# Loss weights
+loss_params:
+  TMA_epoch: 4          # TMA starts at epoch 4
+  lambda_mel: 10.0
+  lambda_gen: 1.0
+  lambda_slm: 1.0
+  lambda_mag: 0.2
+  lambda_entropy: 50.0
+  lambda_f0: 1.0
+  # Note: No lambda_style since we use learnable embeddings
+
+# Optimizer
+optimizer_params:
+  lr: 1e-5
+  min_lr: 1e-5
+  decay_epochs: 6
+  pct_start: 0.33

Configs/config_codec_v2.yml

@@ -0,0 +1,127 @@
+# config_codec_v2.yml
+# ==============================================================================
+# Configuration for TTS Codec V2 Training
+# 
+# Key features:
+#   - Style is tokenized from noised mel spectrogram
+#   - Style tokens are very coarse temporal (~2-8 Hz)
+#   - Content tokens from prosody + text (separate from style)
+# ==============================================================================
+
+log_dir: "codec_v2_training"
+save_freq: 2
+log_interval: 10
+device: "cuda"
+epochs_codec: 100
+batch_size: 16
+max_len: 8000
+
+# Pretrained models
+pretrained_model: "again_wavlm/epoch_flow_00108.pth"  # For text encoder
+pretrained_codec: ""  # Path to resume codec training
+
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 4096
+    win_length: 2400
+    hop_length: 600
+
+model_params:
+  multispeaker: false
+
+  # Base model dimensions
+  dim_in: 64 
+  hidden_dim: 512
+  max_conv_dim: 512
+  n_layer: 3
+  n_mels: 80  # 80 mel bands for style encoder input
+
+  n_token: 178
+  max_dur: 75
+  style_dim: 64
+  style_dim_ac: 128
+  dropout: 0.1
+
+  # ==========================================================================
+  # CODEC V2 SPECIFIC PARAMETERS
+  # ==========================================================================
+  
+  # Style encoder settings
+  # style_compression=16 at hop_length=600 (44.1kHz) = ~4.6 Hz style tokens
+  # style_compression=32 = ~2.3 Hz style tokens (more coarse)
+  style_compression: 16
+  style_dim_codec: 64
+  style_fsq_levels: [8, 8, 8, 8]  # 4096 style codes
+  
+  # Content encoder settings  
+  # content_strides=[2] = 2x compression
+  content_strides: [2]
+  content_fsq_levels: [8, 8, 8, 8, 8, 8]  # 262144 content codes
+  
+  # Codec latent dimensions
+  codec_latent_dim: 256
+  codec_hidden_dim: 512
+
+  # Decoder config (same as ringformer)
+  decoder: 
+    type: 'ringformer'
+    resblock_kernel_sizes: [3, 7, 11]
+    upsample_rates: [4, 6, 10]  # Will be auto-computed based on compression
+    upsample_initial_channel: 512
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    upsample_kernel_sizes: [24, 20]
+    gen_istft_n_fft: 30
+    gen_istft_hop_size: 5
+      
+  # SLM for WavLM loss
+  slm:
+    model: 'microsoft/wavlm-base-plus'
+    sr: 16000
+    hidden: 768
+    nlayers: 13
+    initial_channel: 64
+  
+  # Diffusion config (not used in codec but needed for model building)
+  diffusion:
+    embedding_mask_proba: 0.2
+    transformer:
+      num_layers: 4
+      num_heads: 4
+      head_features: 32
+      multiplier: 4
+    dist:
+      sigma_data: 0.5
+      mean_data: 0.1
+      acoustic: 0.46
+      prosodic: 0.46
+      pitch: 0.46
+      acoustic_mean: 0.4
+      prosodic_mean: 0.4
+      pitch_mean: 0.4
+      estimate_sigma_data: true
+      mean: -0.0
+      std: 0.3
+
+optimizer_params:
+  lr: 1.0e-4
+  pct_start: 0.0
+  
+loss_params:
+  lambda_mel: 45.0           # Mel reconstruction loss
+  lambda_gen: 1.0            # Generator adversarial loss
+  lambda_slm: 1.0            # SLM (WavLM) feature matching loss
+  lambda_mag: 1.0            # Magnitude/phase loss
+  lambda_style_entropy: 1.0  # Style FSQ entropy regularization
+  lambda_content_entropy: 1.0 # Content FSQ entropy regularization
+  TMA_epoch: 5               # Epoch to start adversarial training
+  
+  # Style noise schedule (for robustness)
+  style_noise_start: 0.5     # Initial noise level for style encoder
+  style_noise_end: 0.1       # Final noise level (decays over epochs)

Configs/config_codec_v3.yml

@@ -0,0 +1,131 @@
+# config_codec_v3.yml
+# ==============================================================================
+# Config for TTS Codec V2 with COARSE STYLE TOKENS
+#
+# Style is now [B, 8, T//32] - very coarse temporal tokens from noised mel
+# NOT a continuous vector anymore!
+# ==============================================================================
+
+log_dir: "codec_v3_training"
+save_freq: 2
+log_interval: 10
+device: "cuda"
+epochs_codec: 100
+batch_size: 10
+max_len: 8000
+
+# Pretrained model for text encoder (optional)
+pretrained_model: ""
+pretrained_codec: ""  # Set to load existing codec checkpoint
+
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 4096
+    win_length: 2400
+    hop_length: 600
+
+model_params:
+  multispeaker: false
+  dim_in: 64
+  hidden_dim: 512
+  max_conv_dim: 512
+  n_layer: 3
+  n_mels: 40  # Use 40 mels for style encoder
+
+  n_token: 178
+  max_dur: 75
+  
+  # Style is now COARSE TOKENS, not a continuous vector!
+  # These are codec-specific style settings
+  style_dim: 64           # Base style dim (for old models)
+  style_dim_ac: 128       # Acoustic style (for old models)
+  
+  # ==========================================================================
+  # CODEC V2 STYLE - Temporal style from mel (like original StyleEncoder but temporal)
+  # ==========================================================================
+  # Style tokens: [B, style_dim_codec, T // style_compression]
+  # e.g., [B, 128, T//16] -> temporal style matching original style_dim_ac
+  
+  style_dim_codec: 128          # Match style_dim_ac from original StyleEncoder
+  style_compression: 16         # 16x compression -> ~4.6 Hz at 44.1kHz/600hop
+  
+  # Style FSQ: Increased capacity significantly
+  # [8, 8, 8, 8, 5] -> ~16k codes, 5 dimensions
+  style_fsq_levels: [4, 4, 4, 4, 4]
+  
+  # Noise added to mel during style encoding (for robustness)
+  style_noise_min: 0.0    # Minimum noise level
+  style_noise_max: 0.0    # Maximum noise level (decays during training)
+  
+  # ==========================================================================
+  # CODEC V2 CONTENT - Fine-grained content from pitch/energy/text
+  # ==========================================================================
+  # Content tokens: [B, latent_dim, T // content_compression]
+  # e.g., [B, 256, T//2] means 256-dim content at ~36 Hz (fine-grained!)
+  
+  codec_latent_dim: 256         # Content latent dimension
+  codec_hidden_dim: 512         # Decoder hidden dimension
+  content_strides: [2]          # 2x compression -> ~36 Hz at 44.1kHz/600hop
+  
+  # Content FSQ: 4^5 = 1024 content codes (smaller for 500 samples)
+  # Fine temporal resolution with reasonable codebook size
+  content_fsq_levels: [4, 4, 4, 4, 4]
+  
+  # ==========================================================================
+  # DECODER
+  # ==========================================================================
+  decoder:
+    type: 'ringformer'
+    resblock_kernel_sizes: [3, 7, 11]
+    resblock_dilation_sizes: [[1, 3, 5], [1, 3, 5], [1, 3, 5]]
+    upsample_rates: [4, 6, 10]        # 4*6*10 = 240 (for 2x content compression)
+    upsample_kernel_sizes: [8, 12, 20]
+    upsample_initial_channel: 512
+    gen_istft_n_fft: 30
+    gen_istft_hop_size: 5
+    
+  # ==========================================================================
+  # SLM (Speech Language Model) for discriminator
+  # ==========================================================================
+  slm:
+    model: 'microsoft/wavlm-base-plus'
+    sr: 16000
+    hidden: 768
+    nlayers: 13
+    initial_channel: 64
+
+
+
+# ==========================================================================
+# LOSS WEIGHTS
+# ==========================================================================
+loss_params:
+  # Reconstruction losses
+  lambda_mel: 10.0          # Multi-resolution STFT loss
+  lambda_mag: 1.0           # Magnitude/phase loss for iSTFT
+  
+  # Adversarial losses
+  lambda_gen: 1.0           # Generator adversarial loss
+  lambda_slm: 1.0           # WavLM feature matching loss
+  TMA_epoch: 20              # Start adversarial training at epoch 5
+  
+  # F0 prediction loss (decoder predicts F0 from content tokens)
+  lambda_f0: 10.0           # F0 reconstruction loss
+  
+  # FSQ entropy losses (encourage codebook utilization)
+  lambda_style_fsq: 1.0     # Style FSQ entropy regularization
+  lambda_content_fsq: 1.0   # Content FSQ entropy regularization
+
+# ==========================================================================
+# OPTIMIZER
+# ==========================================================================
+optimizer_params:
+  lr: 0.0001
+  pct_start: 0.0

Configs/config_fsq_flow_convnext.yml

@@ -0,0 +1,78 @@
+# config_fsq_flow_convnext.yml
+# ==============================================================================
+# Configuration for FSQ Flow Predictor with ConvNeXt V2 Text Encoder
+# ==============================================================================
+
+log_dir: "fsq_flow_convnext_training"
+
+# Training hyperparameters
+batch_size: 32
+epochs: 100
+lr: 7e-5
+warmup_steps: 500
+flow_weight: 1.0          # Weight for flow matching loss (boost to speed up convergence)
+ctc_weight: 0.1           # Weight for CTC auxiliary loss
+duration_pred_weight: 1.0 # Weight for duration prediction (reduced to avoid crowding flow gradients)
+cfg_dropout: 0.2          # CFG training dropout prob (20% full cond drop + 20% spk/lang drop)
+dur_sampling_prob: 0.1    # Scheduled duration sampling: use predicted (rescaled) durations ~10% of steps
+
+# Logging and saving
+save_freq: 2      # Save every N epochs
+log_interval: 20  # Log every N batches
+val_interval: 30000 # Validate every N iterations
+
+# Data parameters
+data_params:
+  train_data: "../my_tts_dataset_v2/train_list_filtered_german.txt"
+  val_data: "../my_tts_dataset_v2/val_list_filtered_german.txt"
+  root_path: "../my_tts_dataset_v2"
+  min_length: 50
+  n_mels: 100
+
+# Dump directory for preprocessed features
+dump_dir: "./dump_100"
+
+# Alignment and feature paths
+alignment_train: "../data_preparation_v2/subsets/alignments_train.safetensors"
+alignment_val: "../data_preparation_v2/subsets/alignments_val.safetensors"
+stats_path: "dump_100/speaker_stats.json"
+
+# Precomputed z_continuous directories (from extract_fsq_tokens_for_flow.py)
+z_continuous_train: "precomputed_tokens/train/z_continuous"
+z_continuous_val: "precomputed_tokens/val/z_continuous"
+
+# Pretrained models
+pretrained_model: "fsq_flow_convnext_training/model_epoch_20.pt"  # Path to resume training
+pretrained_codec: "mel_speaker_codec_training/epoch_mel_speaker_codec_00002.pth"  # For validation decoding
+
+# Model parameters
+model_params:
+  # Text Encoder (ConvNeXt V2)
+  vocab_size: 183
+  text_dim: 512
+  text_num_layers: 4
+  text_kernel_size: 7
+  
+  # Language & Speaker
+  num_languages: 10
+  language_dim: 64
+  n_speakers: 11
+  speaker_dim: 128
+  
+  # FSQ Configuration (must match codec)
+  fsq_levels: [4, 4, 4, 4, 4, 4]
+  
+  # Flow Matching Model (DiT)
+  hidden_dim: 512
+  flow_num_layers: 6
+  flow_num_heads: 8
+  
+  # CTC
+  n_phonemes: 176
+
+  # Duration Predictor (Stochastic Flow-based)
+  duration_n_flows: 4
+  duration_kernel_size: 3
+  
+  # Downsampling (100Hz frames -> 50Hz codec tokens)
+  codec_stride: 2

Configs/config_ft.yml

@@ -0,0 +1,111 @@
+log_dir: "Models/LJSpeech"
+save_freq: 5
+log_interval: 10
+device: "cuda"
+epochs: 50 # number of finetuning epoch (1 hour of data)
+batch_size: 8
+max_len: 400 # maximum number of frames
+pretrained_model: "Models/LibriTTS/epochs_2nd_00020.pth"
+second_stage_load_pretrained: true # set to true if the pre-trained model is for 2nd stage
+load_only_params: true # set to true if do not want to load epoch numbers and optimizer parameters
+
+F0_path: "Utils/JDC/bst.t7"
+ASR_config: "Utils/ASR/config.yml"
+ASR_path: "Utils/ASR/epoch_00080.pth"
+PLBERT_dir: 'Utils/PLBERT/'
+
+data_params:
+  train_data: "Data/train_list.txt"
+  val_data: "Data/val_list.txt"
+  root_path: "/local/LJSpeech-1.1/wavs"
+  OOD_data: "Data/OOD_texts.txt"
+  min_length: 50 # sample until texts with this size are obtained for OOD texts
+
+preprocess_params:
+  sr: 24000
+  spect_params:
+    n_fft: 2048
+    win_length: 1200
+    hop_length: 300
+
+model_params:
+  multispeaker: true
+
+  dim_in: 64 
+  hidden_dim: 512
+  max_conv_dim: 512
+  n_layer: 3
+  n_mels: 80
+
+  n_token: 178 # number of phoneme tokens
+  max_dur: 50 # maximum duration of a single phoneme
+  style_dim: 128 # style vector size
+  
+  dropout: 0.2
+
+  # config for decoder
+  decoder: 
+      type: 'hifigan' # either hifigan or istftnet
+      resblock_kernel_sizes: [3,7,11]
+      upsample_rates :  [10,5,3,2]
+      upsample_initial_channel: 512
+      resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]]
+      upsample_kernel_sizes: [20,10,6,4]
+      
+  # speech language model config
+  slm:
+      model: 'microsoft/wavlm-base-plus'
+      sr: 16000 # sampling rate of SLM
+      hidden: 768 # hidden size of SLM
+      nlayers: 13 # number of layers of SLM
+      initial_channel: 64 # initial channels of SLM discriminator head
+  
+  # style diffusion model config
+  diffusion:
+    embedding_mask_proba: 0.1
+    # transformer config
+    transformer:
+      num_layers: 3
+      num_heads: 8
+      head_features: 64
+      multiplier: 2
+
+    # diffusion distribution config
+    dist:
+      sigma_data: 0.2 # placeholder for estimate_sigma_data set to false
+      estimate_sigma_data: true # estimate sigma_data from the current batch if set to true
+      mean: -3.0
+      std: 1.0
+  
+loss_params:
+    lambda_mel: 5. # mel reconstruction loss
+    lambda_gen: 1. # generator loss
+    lambda_slm: 1. # slm feature matching loss
+    
+    lambda_mono: 1. # monotonic alignment loss (TMA)
+    lambda_s2s: 1. # sequence-to-sequence loss (TMA)
+
+    lambda_F0: 1. # F0 reconstruction loss
+    lambda_norm: 1. # norm reconstruction loss
+    lambda_dur: 1. # duration loss
+    lambda_ce: 20. # duration predictor probability output CE loss
+    lambda_sty: 1. # style reconstruction loss
+    lambda_diff: 1. # score matching loss
+    
+    diff_epoch: 10 # style diffusion starting epoch
+    joint_epoch: 30 # joint training starting epoch
+
+optimizer_params:
+  lr: 0.0001 # general learning rate
+  bert_lr: 0.00001 # learning rate for PLBERT
+  ft_lr: 0.0001 # learning rate for acoustic modules
+  
+slmadv_params:
+  min_len: 400 # minimum length of samples
+  max_len: 500 # maximum length of samples
+  batch_percentage: 0.5 # to prevent out of memory, only use half of the original batch size
+  iter: 10 # update the discriminator every this iterations of generator update
+  thresh: 5 # gradient norm above which the gradient is scaled
+  scale: 0.01 # gradient scaling factor for predictors from SLM discriminators
+  sig: 1.5 # sigma for differentiable duration modeling
+  

Configs/config_gemma_audio.yml

@@ -0,0 +1,74 @@
+# config_gemma_audio.yml
+# ==============================================================================
+# Configuration for Gemma Audio Token LM Training
+#
+# This trains Gemma to predict audio tokens autoregressively with:
+# - Text context: sliding window [prev2, prev1, CURRENT, next1]
+# - Audio context: prev2_tokens, prev1_tokens
+# ==============================================================================
+
+# Output directory
+log_dir: "gemma_audio_lm"
+
+# ==============================================================================
+# Training Parameters
+# ==============================================================================
+batch_size: 32                   # Per-GPU batch size (increased for high VRAM)
+gradient_accumulation_steps: 1   # No accumulation needed
+epochs: 50
+save_freq: 2                     # Save every N epochs
+log_interval: 50                 # Log every N steps
+
+# Optimization
+learning_rate: 1.0e-4            # Base LR for audio components
+warmup_steps: 1000               # Linear warmup steps
+max_grad_norm: 1.0               # Gradient clipping
+
+# Sequence lengths
+max_token_len: 512               # Maximum audio token sequence length
+
+# Resume training
+resume_checkpoint: ""            # Path to checkpoint to resume from
+
+# ==============================================================================
+# Data Parameters
+# ==============================================================================
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  token_dir: "./precomputed_tokens"              # Directory with extracted audio tokens
+  alignment_path: "align_outputs/alignments_train.safetensors"
+  alignment_path_val: "align_outputs/alignments_val.safetensors"
+
+# ==============================================================================
+# Model Parameters
+# ==============================================================================
+model_params:
+  # Gemma backbone
+  gemma_model: "google/gemma-3-4b-it"   # Options: gemma-3-1b-it, gemma-2-2b-it, etc.
+  
+  # Audio vocabulary (must match your codec FSQ levels)
+  # FSQ [4]*6 = 4096 codebook size
+  # FSQ [4]*8 = 65536 codebook size
+  audio_vocab_size: 4096
+  
+  # LoRA configuration for efficient fine-tuning
+  use_lora: true
+  lora_r: 16                      # LoRA rank
+  lora_alpha: 32                  # LoRA alpha
+  lora_dropout: 0.05              # LoRA dropout
+  
+  # Audio context encoder
+  audio_context_layers: 4         # Number of transformer layers for audio context
+
+# ==============================================================================
+# Preprocessing Parameters (from your existing config)
+# ==============================================================================
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 4096
+    win_length: 2400
+    hop_length: 600
+    n_mels: 128

Configs/config_libritts.yml

@@ -0,0 +1,113 @@
+log_dir: "Models/LibriTTS"
+first_stage_path: "first_stage.pth"
+save_freq: 1
+log_interval: 10
+device: "cuda"
+epochs_1st: 50 # number of epochs for first stage training (pre-training)
+epochs_2nd: 30 # number of peochs for second stage training (joint training)
+batch_size: 16
+max_len: 300 # maximum number of frames
+pretrained_model: ""
+second_stage_load_pretrained: true # set to true if the pre-trained model is for 2nd stage
+load_only_params: false # set to true if do not want to load epoch numbers and optimizer parameters
+
+F0_path: "Utils/JDC/bst.t7"
+ASR_config: "Utils/ASR/config.yml"
+ASR_path: "Utils/ASR/epoch_00080.pth"
+PLBERT_dir: 'Utils/PLBERT/'
+
+data_params:
+  train_data: "Data/train_list.txt"
+  val_data: "Data/val_list.txt"
+  root_path: ""
+  OOD_data: "Data/OOD_texts.txt"
+  min_length: 50 # sample until texts with this size are obtained for OOD texts
+
+preprocess_params:
+  sr: 24000
+  spect_params:
+    n_fft: 2048
+    win_length: 1200
+    hop_length: 300
+
+model_params:
+  multispeaker: true
+
+  dim_in: 64 
+  hidden_dim: 512
+  max_conv_dim: 512
+  n_layer: 3
+  n_mels: 80
+
+  n_token: 178 # number of phoneme tokens
+  max_dur: 50 # maximum duration of a single phoneme
+  style_dim: 128 # style vector size
+  
+  dropout: 0.2
+
+  # config for decoder
+  decoder: 
+      type: 'hifigan' # either hifigan or istftnet
+      resblock_kernel_sizes: [3,7,11]
+      upsample_rates :  [10,5,3,2]
+      upsample_initial_channel: 512
+      resblock_dilation_sizes: [[1,3,5], [1,3,5], [1,3,5]]
+      upsample_kernel_sizes: [20,10,6,4]
+      
+  # speech language model config
+  slm:
+      model: 'microsoft/wavlm-base-plus'
+      sr: 16000 # sampling rate of SLM
+      hidden: 768 # hidden size of SLM
+      nlayers: 13 # number of layers of SLM
+      initial_channel: 64 # initial channels of SLM discriminator head
+  
+  # style diffusion model config
+  diffusion:
+    embedding_mask_proba: 0.1
+    # transformer config
+    transformer:
+      num_layers: 3
+      num_heads: 8
+      head_features: 64
+      multiplier: 2
+
+    # diffusion distribution config
+    dist:
+      sigma_data: 0.2 # placeholder for estimate_sigma_data set to false
+      estimate_sigma_data: true # estimate sigma_data from the current batch if set to true
+      mean: -3.0
+      std: 1.0
+  
+loss_params:
+    lambda_mel: 5. # mel reconstruction loss
+    lambda_gen: 1. # generator loss
+    lambda_slm: 1. # slm feature matching loss
+    
+    lambda_mono: 1. # monotonic alignment loss (1st stage, TMA)
+    lambda_s2s: 1. # sequence-to-sequence loss (1st stage, TMA)
+    TMA_epoch: 5 # TMA starting epoch (1st stage)
+
+    lambda_F0: 1. # F0 reconstruction loss (2nd stage)
+    lambda_norm: 1. # norm reconstruction loss (2nd stage)
+    lambda_dur: 1. # duration loss (2nd stage)
+    lambda_ce: 20. # duration predictor probability output CE loss (2nd stage)
+    lambda_sty: 1. # style reconstruction loss (2nd stage)
+    lambda_diff: 1. # score matching loss (2nd stage)
+    
+    diff_epoch: 10 # style diffusion starting epoch (2nd stage)
+    joint_epoch: 15 # joint training starting epoch (2nd stage)
+
+optimizer_params:
+  lr: 0.0001 # general learning rate
+  bert_lr: 0.00001 # learning rate for PLBERT
+  ft_lr: 0.00001 # learning rate for acoustic modules
+  
+slmadv_params:
+  min_len: 400 # minimum length of samples
+  max_len: 500 # maximum length of samples
+  batch_percentage: 0.5 # to prevent out of memory, only use half of the original batch size
+  iter: 20 # update the discriminator every this iterations of generator update
+  thresh: 5 # gradient norm above which the gradient is scaled
+  scale: 0.01 # gradient scaling factor for predictors from SLM discriminators
+  sig: 1.5 # sigma for differentiable duration modeling

Configs/config_predictors_speaker.yml

@@ -0,0 +1,61 @@
+# config_predictors_speaker.yml
+# ==============================================================================
+# Configuration for Duration and Token Predictors with Speaker ID
+#
+# ARCHITECTURE:
+#   - Duration Predictor: Text + Speaker ID -> Durations (Flow-based)
+#   - Token Predictor: Text + Durations + Speaker ID -> Codec Tokens
+#
+# No style vectors - only speaker embeddings for conditioning
+# Codec uses stride [2] for 2x temporal compression
+# ==============================================================================
+
+log_dir: "predictors_speaker_training"
+save_freq: 2
+log_interval: 10
+
+# Training
+epochs: 100
+batch_size: 32
+lr: 1e-4
+
+# Pretrained model (optional)
+pretrained_model: ""
+
+# Precomputed codec tokens (from extract_codec_tokens.py)
+# These are the target tokens for the token predictor
+codec_tokens_train: "precomputed_tokens/codec_tokens_train.safetensors"
+codec_tokens_val: "precomputed_tokens/codec_tokens_val.safetensors"
+
+# Data
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 4096
+    win_length: 2400
+    hop_length: 600
+
+# Model parameters
+model_params:
+  # Vocabulary
+  n_token: 178
+  
+  # Speaker settings
+  n_speakers: 100
+  speaker_emb_dim: 256
+  
+  # Model dimensions
+  hidden_dim: 512
+  
+  # Codec settings (must match your trained codec in config_codec_hybrid.yml)
+  # codec_strides: [2, 2] means 4x temporal compression
+  # But if you want stride [2] only, use codec_stride: 2
+  codec_stride: 2  # Product of [2]
+  # FSQ codebook size: product of levels [4, 4, 4, 4, 4] = 4^5 = 1024
+  codebook_size: 1024

Configs/config_predictors_speaker_diffusion.yml

@@ -0,0 +1,60 @@
+# config_predictors_speaker_diffusion.yml
+# ==============================================================================
+# Configuration for Token Predictor with Flow Matching + DiT
+# ==============================================================================
+
+log_dir: "predictors_speaker_diffusion_training"
+
+# Training params
+batch_size: 16
+epochs: 100
+save_freq: 2
+log_interval: 10
+lr: 1e-4
+
+# Data params
+data_params:
+  train_data: "./Data/train_list.txt"
+  val_data: "./Data/val_list.txt"
+  root_path: "/path/to/audio/root"
+  min_length: 50
+
+# Model params
+model_params:
+  n_speakers: 100
+  speaker_emb_dim: 256
+  hidden_dim: 512
+  n_token: 178
+  codec_stride: 2
+  
+  # FSQ levels for codec tokens
+  fsq_levels: [4, 4, 4, 4, 4]
+  
+  # Gemma 1B hidden dimension
+  gemma_dim: 1152
+
+# Flow Matching params
+flow_params:
+  # DiT encoder layers
+  n_layers_enc: 4
+  
+  # Flow matching velocity network layers
+  n_layers_flow: 6
+  
+  # ODE solver steps for inference
+  flow_steps: 20
+  
+  # Classifier-free guidance scale
+  cfg_scale: 1.5
+
+# Precomputed data paths
+codec_tokens_train: "precomputed_tokens/codec_tokens_train.safetensors"
+codec_tokens_val: "precomputed_tokens/codec_tokens_val.safetensors"
+
+# Gemma embeddings (precomputed from raw_texts)
+gemma_embeddings_train: "precomputed_tokens/gemma_embeddings_train.safetensors"
+gemma_embeddings_val: "precomputed_tokens/gemma_embeddings_val.safetensors"
+
+# Pretrained models
+pretrained_model: ""  # Resume from checkpoint
+pretrained_hybrid_codec: "hybrid_codec_training/epoch_hybrid_codec_00017.pth"

Configs/config_predictors_speaker_flow.yml

@@ -0,0 +1,89 @@
+# config_predictors_speaker_flow.yml
+# ==============================================================================
+# Configuration for Duration and Token Predictors with Speaker ID
+# Token Predictor uses Flow Matching (CFM) instead of classification
+#
+# ARCHITECTURE:
+#   - Duration Predictor: Text + Speaker ID -> Durations (Flow-based SDP)
+#   - Token Predictor: Text + Durations + Speaker ID -> FSQ Embeddings (CFM)
+#
+# Key Features:
+#   - Continuous token prediction via Conditional Flow Matching
+#   - Predicts FSQ embeddings directly (not discrete token IDs)
+#   - At inference, quantizes to nearest FSQ codes
+#   - More stable training, smoother outputs
+# ==============================================================================
+
+log_dir: "predictors_speaker_flow_training"
+save_freq: 2
+log_interval: 10
+
+# Training
+epochs: 100
+batch_size: 32
+lr: 1e-4
+
+# Gradient clipping (recommended for flow matching)
+grad_clip: 1.0
+
+# Loss weights
+loss_weight_dur: 1.0
+loss_weight_token: 100.0  # CFM loss is typically smaller scale
+
+# Pretrained model (optional)
+pretrained_model: ""
+
+# Precomputed data
+# Option 1: FSQ embeddings (preferred - continuous targets)
+fsq_embeddings_train: "precomputed_tokens/fsq_embeddings_train.safetensors"
+fsq_embeddings_val: "precomputed_tokens/fsq_embeddings_val.safetensors"
+
+# Option 2: Codec tokens (fallback - will convert to embeddings)
+codec_tokens_train: "precomputed_tokens/codec_tokens_train.safetensors"
+codec_tokens_val: "precomputed_tokens/codec_tokens_val.safetensors"
+
+# Data
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 4096
+    win_length: 2400
+    hop_length: 600
+
+# Model parameters
+model_params:
+  # Vocabulary
+  n_token: 178
+  
+  # Speaker settings
+  n_speakers: 100
+  speaker_emb_dim: 256
+  
+  # Model dimensions
+  hidden_dim: 512
+  
+  # Codec settings
+  codec_stride: 2
+  
+  # FSQ settings (must match your trained codec)
+  fsq_levels: [4, 4, 4, 4, 4]  # 4^5 = 1024 codes
+
+# CFM (Conditional Flow Matching) parameters
+cfm_params:
+  # Hidden dimension for CFM estimator
+  hidden_channels: 512
+  
+  # Number of residual blocks in CFM
+  n_layers: 6
+  
+  # Dropout rate
+  dropout: 0.1
+  
+  # ODE solver steps at inference
+  solver_steps: 20

Configs/config_predictors_speaker_flow_context.yml

@@ -0,0 +1,103 @@
+# config_predictors_speaker_flow_context.yml
+# ==============================================================================
+# Configuration for Duration and Token Predictors with Speaker ID + Context
+# Token Predictor uses Flow Matching (CFM) approach
+#
+# ARCHITECTURE:
+#   - Duration Predictor: Text + Context + Speaker ID -> Durations (Flow-based SDP)
+#   - Token Predictor: Text + Context + Durations + Speaker ID -> FSQ Embeddings (CFM)
+#
+# Key Features:
+#   - Context-aware predictions for smooth sliding window inference
+#   - Cross-attention to neighboring sentences (prev/next)
+#   - Exclusion mask prevents attending to current sentence in context
+#   - This ensures sharp predictions while maintaining context awareness
+# ==============================================================================
+
+log_dir: "predictors_context_training"
+save_freq: 2
+log_interval: 10
+
+# Training
+epochs: 100
+batch_size: 32
+lr: 1e-4
+
+# Gradient clipping (recommended for flow matching)
+grad_clip: 1.0
+
+# Loss weights
+loss_weight_dur: 1.0
+loss_weight_token: 100.0  # CFM loss is typically smaller scale
+
+# Pretrained model (optional)
+pretrained_model: "predictors_context_training/epoch_00100.pth"
+
+# Precomputed data
+# Option 1: FSQ embeddings (preferred - continuous targets)
+fsq_embeddings_train: "precomputed_tokens/fsq_embeddings_train.safetensors"
+fsq_embeddings_val: "precomputed_tokens/fsq_embeddings_val.safetensors"
+
+# Option 2: Codec tokens (fallback - will convert to embeddings)
+codec_tokens_train: "precomputed_tokens/codec_tokens_train.safetensors"
+codec_tokens_val: "precomputed_tokens/codec_tokens_val.safetensors"
+
+# Pretrained hybrid codec (for validation audio)
+pretrained_hybrid_codec: "hybrid_codec_training/epoch_mel_speaker_codec_00002.pth"
+
+# Data
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 4096
+    win_length: 2400
+    hop_length: 600
+
+# Model parameters
+model_params:
+  # Vocabulary
+  n_token: 178
+  
+  # Speaker settings
+  n_speakers: 100
+  speaker_emb_dim: 256
+  
+  # Model dimensions
+  hidden_dim: 512
+  
+  # Codec settings
+  codec_stride: 2
+  
+  # FSQ settings (must match your trained codec)
+  fsq_levels: [4, 4, 4, 4, 4]  # 4^5 = 1024 codes
+
+# Context conditioning parameters
+context_params:
+  # Number of transformer layers in context encoder
+  n_context_layers: 2
+  
+  # Number of attention heads in context encoder
+  n_context_heads: 4
+  
+  # Enable cross-attention from current tokens to context
+  use_context_cross_attn: true
+
+# CFM (Conditional Flow Matching) parameters
+cfm_params:
+  # Hidden dimension for CFM estimator
+  hidden_channels: 512
+  
+  # Number of residual blocks in CFM
+  n_layers: 6
+  
+  # Dropout rate
+  dropout: 0.1
+  
+  # ODE solver steps at inference
+  solver_steps: 20

Configs/config_predictors_speaker_flow_context_temporal.yml

@@ -0,0 +1,88 @@
+# config_predictors_speaker_flow_context_temporal.yml
+# ==============================================================================
+# Configuration for Duration and Token Predictors with Speaker ID + Context + Temporal Codec
+# Token Predictor uses Flow Matching (CFM) approach
+# ==============================================================================
+
+log_dir: "predictors_speaker_flow_context_temporal_training"
+save_freq: 2
+log_interval: 10
+
+# Training
+epochs: 100
+batch_size: 32
+lr: 1e-4
+
+# Gradient clipping (recommended for flow matching)
+grad_clip: 2.0
+
+# Loss weights
+loss_weight_dur: 1.0
+loss_weight_token: 100.0  # CFM loss is typically smaller scale
+
+# Pretrained model (optional)
+pretrained_model: ""
+
+# Precomputed data
+codec_tokens_train: "precomputed_tokens_temporal/codec_tokens_temporal_train.safetensors"
+# codec_tokens_val: "precomputed_tokens_temporal/codec_tokens_temporal_val.safetensors" # Optional
+
+# Pretrained hybrid temporal codec (for validation audio generation)
+pretrained_hybrid_temporal_codec: "hybrid_temporal_codec_training/epoch_hybrid_temporal_codec_00031.pth"
+
+# Data
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  min_length: 50
+
+preprocess_params:
+  sr: 44100
+  spect_params:
+    n_fft: 2048
+    win_length: 1764
+    hop_length: 441 
+
+# Model parameters
+model_params:
+  # Vocabulary
+  n_token: 176
+  
+  # Speaker settings
+  n_speakers: 4
+  speaker_emb_dim: 256
+  
+  # Model dimensions
+  hidden_dim: 512
+  
+  # Codec settings
+  codec_stride: 2
+  
+  # FSQ settings (must match your trained codec)
+  fsq_levels: [4, 4, 4, 4, 4]  # 4^5 = 1024 codes
+
+# Context conditioning parameters
+context_params:
+  # Number of transformer layers in context encoder
+  n_context_layers: 2
+  
+  # Number of attention heads in context encoder
+  n_context_heads: 4
+  
+  # Enable cross-attention from current tokens to context
+  use_context_cross_attn: true
+
+# CFM (Conditional Flow Matching) parameters
+cfm_params:
+  # Hidden dimension for CFM estimator
+  hidden_channels: 512
+  
+  # Number of residual blocks in CFM
+  n_layers: 3
+  
+  # Dropout rate
+  dropout: 0.1
+  
+  # Solver steps for inference
+  solver_steps: 20

Configs/config_predictors_streaming_dit.yml

@@ -0,0 +1,100 @@
+# Configuration for Streaming-Native Predictors with DiT
+# ==============================================================================
+# Trains duration and token predictors with:
+# - Block-structured attention: 3 blocks receptive field
+#   * Current block (1)
+#   * 2-block lookback
+#   * 1-block lookahead
+# - Block-based processing during training and inference
+# - DiT backbone instead of flow-based models
+# - Flow Matching for token prediction
+# ==============================================================================
+
+log_dir: "logs/predictors_streaming_dit"
+
+# Training
+epochs: 100
+batch_size: 16  # Can be smaller than standard due to block processing
+save_freq: 2
+log_interval: 10
+lr: 1.0e-4
+
+# Data
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  min_length: 50
+
+# Precomputed codec tokens (from extract_codec_tokens_hybrid_temporal.py)
+codec_tokens_train: "precomputed_tokens_temporal/codec_tokens_temporal_train.safetensors"
+
+# Pretrained models
+pretrained_model: ""  # Resume training from checkpoint
+pretrained_hybrid_temporal_codec: "hybrid_temporal_codec_training/epoch_hybrid_temporal_codec_00031.pth"
+
+# Model architecture
+model_params:
+  n_token: 178  # Vocabulary size
+  n_speakers: 100
+  speaker_emb_dim: 256
+  d_model: 512  # Hidden dimension for all components
+  codec_stride: 2  # Codec compression (50Hz tokens)
+  fsq_levels: [4, 4, 4, 4, 4]  # FSQ quantization levels
+
+# Streaming parameters (block-structured attention)
+streaming_params:
+  # Block size: number of tokens per block
+  block_size: 50
+  
+  # Receptive field: 3 blocks total
+  # - Current block (1)
+  # - 2-block lookback
+  # - 1-block lookahead
+  n_lookback_blocks: 2
+  n_lookahead_blocks: 1
+  
+  # Enable block-based training for long sequences
+  use_block_training: true
+
+# DiT architecture
+dit_params:
+  # Text encoder (shared between duration and token predictors)
+  encoder_layers: 4
+  encoder_heads: 8
+  
+  # Duration predictor DiT
+  duration_dit_layers: 6
+  duration_dit_heads: 8
+  
+  # Token predictor DiT (for CFM velocity estimation)
+  token_dit_layers: 6
+  token_dit_heads: 8
+
+# Flow Matching (for token predictor)
+cfm_params:
+  solver_steps: 20  # Number of ODE solver steps
+  sigma_min: 1.0e-4  # Minimum noise level
+
+# Optimizer
+optimizer_params:
+  lr: 1.0e-4
+  betas: [0.9, 0.999]
+  eps: 1.0e-8
+
+# Notes:
+# - BLOCK-STRUCTURED ATTENTION:
+#   * The DiT's receptive field is restricted to 3 blocks
+#   * Current block + 2 lookback + 1 lookahead
+#   * Example: block_size=50 → total receptive field = 4 blocks × 50 = 200 tokens
+#   * Structured as discrete blocks, NOT continuous sliding window
+#   * Memory: O(N × 4B) where B is block_size
+#
+# - block_size controls the granularity
+#   * 50 = each block contains 50 text tokens
+#   * Larger = fewer blocks, more tokens per block
+#   * Smaller = more blocks, finer granularity
+#
+# - use_block_training enables automatic block creation during training
+#   * true = automatically chunk long sequences into blocks with proper masks
+#   * false = process full sequences (may OOM on long texts)

Configs/config_stage2_masked.yml

@@ -0,0 +1,69 @@
+# config_stage2_masked.yml
+# ==============================================================================
+# Configuration for Stage 2 Training: Duration + Masked Token Predictors
+#
+# This config is for training the predictors that work with the
+# HybridTTSCodecVocoderTemporal codec.
+# ==============================================================================
+
+# Output directory
+log_dir: "logs_stage2_masked"
+
+# Training parameters
+batch_size: 16
+epochs: 100
+lr: 1e-4
+save_freq: 5
+log_interval: 20
+
+# Loss weights
+lambda_dur: 1.0
+lambda_token: 1.0
+
+# Data parameters
+data_params:
+  train_data: "Data/train_3_filtered_dur_fixed_speakers_test.txt"
+  val_data: "Data/val_3_filtered_fixed_speakers.txt"
+  root_path: "../tts_model"
+  min_length: 50
+
+# Model parameters
+model_params:
+  n_speakers: 4
+  speaker_emb_dim: 256
+  hidden_dim: 512
+  n_token: 178
+  codec_stride: 2
+  fsq_levels: [4, 4, 4, 4, 4]  # Must match codec training
+
+# Context parameters (for sliding window)
+context_params:
+  n_context_layers: 2
+  n_context_heads: 4
+  use_context_cross_attn: true
+
+# Masked Transformer parameters (for token prediction)
+mask_params:
+  n_layers: 8
+  n_heads: 8
+  dropout: 0.1
+
+# Duration predictor parameters
+duration_params:
+  use_deterministic_head: false  # false = stochastic flow-based
+  n_flows: 4
+
+# Precomputed tokens (from extract_codec_tokens_hybrid_temporal.py)
+codec_tokens_train: "precomputed_tokens_temporal/codec_tokens_temporal_train.safetensors"
+
+# Pretrained models
+pretrained_model: ""  # Stage 2 checkpoint to resume from (optional)
+pretrained_hybrid_temporal_codec: "logs_codec_hybrid_temporal/hybrid_temporal_codec_best.pth"
+
+# Preprocessing (should match codec config)
+preprocess_params:
+  sr: 44100
+  spect_params:
+    hop_length: 441
+    n_fft: 4096
+    win_length: 1764

Data/OOD_texts.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:babec1dc6881a3b10ecb11b5de56fae8a02923f831c5d2e2da819841a86a5ecc
+size 15667506

Data/add.py

@@ -0,0 +1,31 @@
+#!/usr/bin/env python3
+
+INPUT_FILE = "fashion_train.txt"
+OUTPUT_FILE = "fashion.txt"
+
+def main():
+    out_lines = []
+    with open(INPUT_FILE, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.rstrip("\n")
+            if not line.strip():
+                continue
+
+            parts = line.split("|")
+            # Expecting: path | phonemes | speaker_id | text
+            if len(parts) == 4:
+                path, phonemes, speaker_id, text = parts
+                new_line = f"{path}|{phonemes}|{speaker_id}|0|{text}"
+                out_lines.append(new_line)
+            else:
+                # If unexpected format, keep line as-is
+                out_lines.append(line)
+
+    with open(OUTPUT_FILE, "w", encoding="utf-8") as f:
+        for l in out_lines:
+            f.write(l + "\n")
+
+    print(f"✅ Added language column successfully → {OUTPUT_FILE}")
+
+if __name__ == "__main__":
+    main()

Data/check_list.py

@@ -0,0 +1,68 @@
+import os
+import numpy as np
+from scipy.io import wavfile
+
+def is_extreme(wav_path, threshold=95):
+    """
+    Return True if any sample in wav_path exceeds ±threshold% of full-scale.
+    """
+    try:
+        _, data = wavfile.read(wav_path)
+    except Exception as e:
+        print(f"[SKIP]   {wav_path!r}: {e}")
+        # Treat unreadable files as "extreme" so they're dropped
+        return True
+
+    samples = data.flatten().astype(np.float32)
+    dt = data.dtype
+
+    if dt == np.int16:
+        max_val = 32767.0
+    elif dt == np.int32:
+        max_val = 2147483647.0
+    elif dt == np.uint8:
+        samples -= 128.0
+        max_val = 127.0
+    else:
+        # assume float32/64 normalized to [-1,1]
+        max_val = 1.0
+
+    thresh_val = (threshold / 100.0) * max_val
+    return np.any(samples >  thresh_val) or np.any(samples < -thresh_val)
+
+def filter_to_new(input_list,
+                  output_list,
+                  prefix="../",
+                  threshold=95):
+    """
+    Read input_list, drop lines whose wavs exceed ±threshold%, and
+    write the remaining lines to output_list.
+    """
+    with open(input_list, "r", encoding="utf-8") as f:
+        lines = [L for L in f if L.strip()]
+
+    kept = []
+    for L in lines:
+        rel_path = L.split("|", 1)[0].strip()
+        wav_path = os.path.join(prefix, rel_path)
+        if not os.path.isfile(wav_path):
+            print(f"[MISSING] {wav_path}")
+            continue
+        if is_extreme(wav_path, threshold):
+            print(f"[DROP]    {wav_path}")
+        else:
+            kept.append(L)
+
+    with open(output_list, "w", encoding="utf-8") as f:
+        f.writelines(kept)
+
+    print(f"\nKept {len(kept)} of {len(lines)} entries.")
+    print(f"Written filtered list to: {output_list}")
+
+if __name__ == "__main__":
+    filter_to_new(
+        input_list="train_list_filtered.txt",
+        output_list="train_list_no_clipped_90.txt",
+        prefix="../",
+        threshold=90
+    )

Data/dd.py

@@ -0,0 +1,123 @@
+import os
+from phonemizer import backend
+
+# Konfiguracja plików
+INPUT_FILE = 'train_3_filtered_dur.txt'        # Plik wejściowy
+OUTPUT_FILE = 'train_3_filtered_dur_fixed.txt' # Plik wynikowy
+
+# Zestaw znaków, których obecność w fonemach powoduje usunięcie linii
+FORBIDDEN_CHARS = {'{', '}', '(', ')'}
+
+def initialize_phonemizers():
+    """
+    Inicjalizuje słownik backendów espeak dla wymaganych języków.
+    Zwraca słownik { 'language_id': backend_instance }
+    """
+    phonemizers = {}
+    print("Inicjalizacja phonemizerów (może to chwilę potrwać)...")
+    
+    try:
+        # ID 0: Polski
+        phonemizers['0'] = backend.EspeakBackend(
+            language='pl', 
+            preserve_punctuation=True, 
+            with_stress=True
+        )
+        
+        # ID 1: Angielski Brytyjski
+        phonemizers['1'] = backend.EspeakBackend(
+            language='en-gb', 
+            preserve_punctuation=True, 
+            with_stress=True
+        )
+        
+        # ID 2: Hindi
+        phonemizers['2'] = backend.EspeakBackend(
+            language='hi', 
+            preserve_punctuation=True, 
+            with_stress=True
+        )
+        
+        print("Phonemizery zainicjalizowane pomyślnie.")
+        return phonemizers
+
+    except Exception as e:
+        print(f"Krytyczny błąd inicjalizacji phonemizerów: {e}")
+        print("Upewnij się, że masz zainstalowany 'espeak-ng'.")
+        return None
+
+def process_dataset():
+    # 1. Przygotuj phonemizery
+    phonemizers = initialize_phonemizers()
+    if not phonemizers:
+        return
+
+    if not os.path.exists(INPUT_FILE):
+        print(f"Nie znaleziono pliku: {INPUT_FILE}")
+        return
+
+    print(f"Rozpoczynam przetwarzanie pliku {INPUT_FILE}...")
+    
+    processed_count = 0
+    skipped_count = 0
+    
+    with open(INPUT_FILE, 'r', encoding='utf-8') as f_in, \
+         open(OUTPUT_FILE, 'w', encoding='utf-8') as f_out:
+        
+        for line_num, line in enumerate(f_in, 1):
+            line = line.strip()
+            if not line:
+                continue
+                
+            # 2. Split z limitem 4
+            parts = line.split('|', 4)
+            
+            # Sprawdzenie poprawności struktury (musi być 5 kolumn)
+            if len(parts) < 5:
+                print(f"Ostrzeżenie: Linia {line_num} ma nieprawidłowy format (pominięto).")
+                continue
+
+            # Wyciągamy potrzebne dane
+            # parts[1] = phonemes (to podmieniamy i sprawdzamy)
+            lang_id = parts[3]
+            raw_text = parts[4]
+            
+            # Pobieramy odpowiedni phonemizer
+            current_backend = phonemizers.get(lang_id)
+            
+            if current_backend:
+                try:
+                    # Generowanie fonemów
+                    phonemes_list = current_backend.phonemize([raw_text], strip=True)
+                    if phonemes_list:
+                        parts[1] = phonemes_list[0]  # Aktualizacja fonemów
+                except Exception as e:
+                    print(f"Błąd phonemizera w linii {line_num} (ID jęz: {lang_id}): {e}")
+            
+            # --- NOWA LOGIKA FILTROWANIA ---
+            current_phonemes = parts[1]
+            
+            # Sprawdzamy czy w fonemach występuje któryś z zakazanych znaków
+            if any(char in current_phonemes for char in FORBIDDEN_CHARS):
+                skipped_count += 1
+                # Opcjonalnie: odkomentuj linię poniżej, jeśli chcesz widzieć, co jest usuwane
+                # print(f"Pominięto linię {line_num}: znaleziono zakazany znak w fonemach: {current_phonemes}")
+                continue
+
+            # 3. Zapis do nowego pliku (tylko jeśli walidacja przeszła pomyślnie)
+            processed_count += 1
+            new_line = "|".join(parts)
+            f_out.write(new_line + '\n')
+
+            # Logowanie postępu co 100 linii
+            if processed_count > 0 and processed_count % 100 == 0:
+                print(f"Zapisano {processed_count} poprawnych linii...")
+
+    print("-" * 30)
+    print(f"Zakończono!")
+    print(f"Poprawnie przetworzono i zapisano: {processed_count}")
+    print(f"Usunięto (zawierały '{{', '}}', '(', ')'): {skipped_count}")
+    print(f"Wynik zapisano w: {OUTPUT_FILE}")
+
+if __name__ == "__main__":
+    process_dataset()

Data/del_amp.py

@@ -0,0 +1,58 @@
+from pathlib import Path
+import numpy as np
+import soundfile as sf
+
+# USTAWIENIA (statycznie)
+ROOT_DIR = Path("Dataset")  # <- zmień na swoją ścieżkę
+
+# Przedziały (włącznie z granicami)
+MIN_SEC = 1.3
+MAX_SEC = 15.0
+
+AMP_MIN = 0.01  # minimalny peak |x|
+AMP_MAX = 1.0  # maksymalny peak |x|
+
+def iter_wavs(root: Path):
+    for p in root.rglob("*.wav"):
+        if p.is_file():
+            yield p
+
+def should_keep(wav: Path) -> bool:
+    try:
+        data, sr = sf.read(wav, always_2d=True)  # shape: (n_samples, n_channels)
+        n, _ = data.shape
+        if sr <= 0 or n == 0:
+            return False
+
+        # Długość
+        dur = n / float(sr)
+        if not (MIN_SEC <= dur <= MAX_SEC):
+            return False
+
+        # Peak amplitudy w mono
+        mono = data.mean(axis=1).astype(np.float32, copy=False)
+        peak = float(np.max(np.abs(mono)))
+
+        return (AMP_MIN <= peak <= AMP_MAX)
+    except Exception:
+        return False
+
+def main():
+    kept = 0
+    deleted = 0
+    errors = 0
+
+    for wav in iter_wavs(ROOT_DIR):
+        if should_keep(wav):
+            kept += 1
+        else:
+            try:
+                wav.unlink()  # USUŃ plik niespełniający warunków
+                deleted += 1
+            except Exception:
+                errors += 1
+
+    print(f"Zachowane: {kept} | Usunięte: {deleted} | Błędy: {errors}")
+
+if __name__ == "__main__":
+    main()

Data/delete_silence.py

@@ -0,0 +1,22 @@
+#!/usr/bin/env python3
+import os
+import numpy as np
+import soundfile as sf
+
+# Ścieżka do katalogu wavs
+base_dir = 'Dataset/wavs'
+# Próg amplitudy – pliki, które nie mają próbki > threshold, zostaną usunięte
+threshold = 0.07
+
+# Iteruj przez wszystkie podfoldery i pliki
+for root, dirs, files in os.walk(base_dir):
+    for file in files:
+        if not file.lower().endswith('.wav'):
+            continue
+        full_path = os.path.join(root, file).replace('\\', '/')
+        # Wczytaj dane audio (może być wielokanałowe)
+        data, sr = sf.read(full_path)
+        # Jeśli nie ma próbki przekraczającej threshold → usuń
+        if not np.any(np.abs(data) > threshold):
+            print(f"Usuwam: {full_path}")
+            os.remove(full_path)

Data/filet_train.py

@@ -0,0 +1,332 @@
+import os
+import wave
+import numpy as np
+import subprocess
+import tempfile
+
+
+def filter_audio_list(
+    input_path="train_list.txt",
+    output_path="train_list_filtered.txt",
+    min_duration_sec=1.25,
+    max_duration_sec=10.0,
+    min_amplitude_threshold=0.02,
+    max_amplitude_threshold=0.94,
+    ffmpeg_path="ffmpeg",
+):
+    """
+    Filter lines in input_path according to duration and amplitude checks.
+    Removes duplicate paths: if a given .wav path appears more than once in the
+    input file, ALL occurrences are removed from the output.
+    """
+    if not os.path.exists(input_path):
+        print(f"Error: Input file not found at '{input_path}'")
+        return
+
+    kept_lines = []
+    removed_count = 0
+    total_count = 0
+
+    print(f"Starting to process '{input_path}'...")
+    print(f" - Minimum duration: {min_duration_sec}s")
+    print(f" - Maximum duration: {max_duration_sec}s")
+    print(f" - Minimum amplitude: {min_amplitude_threshold}")
+    print(f" - Maximum amplitude: {max_amplitude_threshold}")
+    print(
+        f" - Target format: 44.1kHz, 16-bit PCM, mono "
+        f"(will convert using {ffmpeg_path})"
+    )
+
+    # Read input lines (preserve full line content for output)
+    with open(input_path, "r", encoding="utf-8") as f:
+        raw_lines = f.readlines()
+
+    # Build canonical path for each non-empty line so we can detect duplicates
+    base_dir = os.path.dirname(os.path.abspath(input_path)) or os.getcwd()
+    canon_for_line = [None] * len(raw_lines)
+    canon_to_indices = {}
+
+    for idx, raw in enumerate(raw_lines):
+        ln = raw.strip()
+        if not ln:
+            continue
+        audio_rel = ln.split("|")[0].strip()
+        # Match original behavior of prefixing "../" relative to the input file
+        candidate = os.path.join(base_dir, "..", audio_rel)
+        canonical = os.path.realpath(os.path.abspath(candidate))
+        canon_for_line[idx] = canonical
+        canon_to_indices.setdefault(canonical, []).append(idx)
+
+    # Identify duplicate canonical paths (appear more than once)
+    duplicates_canonical = {p for p, idxs in canon_to_indices.items() if len(idxs) > 1}
+    if duplicates_canonical:
+        print(
+            f"  [Deduplication] Found {len(duplicates_canonical)} duplicate "
+            "path(s). All occurrences will be removed."
+        )
+        for p in sorted(duplicates_canonical):
+            print(f"    - {p} (occurrences={len(canon_to_indices[p])})")
+
+    # Process lines, skipping duplicates early
+    for idx, raw in enumerate(raw_lines):
+        total_count += 1
+        line = raw.strip()
+        if not line:
+            continue
+
+        canonical = canon_for_line[idx]
+        if canonical in duplicates_canonical:
+            # remove duplicate occurrences (do not keep any of them)
+            print(f"  [Removing] Duplicate entry: {line.split('|')[0].strip()}")
+            removed_count += 1
+            continue
+
+        # Use the canonical absolute path for filesystem operations
+        audio_path = canonical
+
+        try:
+            if not os.path.exists(audio_path):
+                print(f"  [Warning] File not found, skipping: {audio_path}")
+                removed_count += 1
+                continue
+
+            # Try to open as WAV; if not readable, convert with ffmpeg
+            try:
+                wav_file = wave.open(audio_path, "rb")
+            except wave.Error:
+                print(
+                    f"  [Converting] Not a readable WAV: converting {audio_path} -> "
+                    "44.1kHz 16-bit mono WAV"
+                )
+                tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")
+                tmp.close()
+                cmd = [
+                    ffmpeg_path,
+                    "-y",
+                    "-i",
+                    audio_path,
+                    "-ar",
+                    "44100",
+                    "-ac",
+                    "1",
+                    "-acodec",
+                    "pcm_s16le",
+                    tmp.name,
+                ]
+                proc = subprocess.run(
+                    cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE
+                )
+                if proc.returncode != 0:
+                    stderr = proc.stderr.decode("utf-8", errors="replace")
+                    print(
+                        f"  [Error] ffmpeg conversion failed for {audio_path}: "
+                        f"{stderr}"
+                    )
+                    removed_count += 1
+                    try:
+                        os.unlink(tmp.name)
+                    except Exception:
+                        pass
+                    continue
+
+                try:
+                    os.replace(tmp.name, audio_path)
+                except Exception as e:
+                    print(
+                        f"  [Error] Could not replace original with converted file: "
+                        f"{e}"
+                    )
+                    removed_count += 1
+                    try:
+                        os.unlink(tmp.name)
+                    except Exception:
+                        pass
+                    continue
+
+                try:
+                    wav_file = wave.open(audio_path, "rb")
+                except wave.Error as e:
+                    print(
+                        f"  [Warning] Still cannot open after conversion: "
+                        f"{audio_path} ({e})"
+                    )
+                    removed_count += 1
+                    continue
+
+            # Read header info then close
+            try:
+                num_frames = wav_file.getnframes()
+                framerate = wav_file.getframerate()
+                sampwidth = wav_file.getsampwidth()
+                n_channels = wav_file.getnchannels()
+            finally:
+                try:
+                    wav_file.close()
+                except Exception:
+                    pass
+
+            # Convert to target format if needed
+            needs_conversion = (
+                framerate != 44100 or sampwidth != 2 or n_channels != 1
+            )
+            if needs_conversion:
+                print(
+                    f"  [Converting] {audio_path} (sr={framerate}, sw={sampwidth}, "
+                    f"ch={n_channels}) -> 44.1kHz 16-bit mono"
+                )
+                tmp = tempfile.NamedTemporaryFile(delete=False, suffix=".wav")
+                tmp.close()
+                cmd = [
+                    ffmpeg_path,
+                    "-y",
+                    "-i",
+                    audio_path,
+                    "-ar",
+                    "44100",
+                    "-ac",
+                    "1",
+                    "-acodec",
+                    "pcm_s16le",
+                    tmp.name,
+                ]
+                proc = subprocess.run(
+                    cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE
+                )
+                if proc.returncode != 0:
+                    stderr = proc.stderr.decode("utf-8", errors="replace")
+                    print(
+                        f"  [Error] ffmpeg conversion failed for {audio_path}: "
+                        f"{stderr}"
+                    )
+                    removed_count += 1
+                    try:
+                        os.unlink(tmp.name)
+                    except Exception:
+                        pass
+                    continue
+
+                try:
+                    os.replace(tmp.name, audio_path)
+                except Exception as e:
+                    print(
+                        f"  [Error] Could not replace original with converted file: "
+                        f"{e}"
+                    )
+                    removed_count += 1
+                    try:
+                        os.unlink(tmp.name)
+                    except Exception:
+                        pass
+                    continue
+
+            # Open the (possibly converted) WAV and perform checks
+            try:
+                with wave.open(audio_path, "rb") as wf:
+                    num_frames = wf.getnframes()
+                    framerate = wf.getframerate()
+                    sampwidth = wf.getsampwidth()
+                    n_channels = wf.getnchannels()
+
+                    # Duration checks
+                    duration = num_frames / float(framerate)
+                    if duration < min_duration_sec:
+                        print(
+                            f"  [Removing] Short duration: {audio_path} "
+                            f"({duration:.2f}s)"
+                        )
+                        removed_count += 1
+                        continue
+
+                    if duration > max_duration_sec:
+                        print(
+                            f"  [Removing] Too long: {audio_path} "
+                            f"({duration:.2f}s > {max_duration_sec}s)"
+                        )
+                        removed_count += 1
+                        continue
+
+                    # Amplitude check (16-bit PCM)
+                    if sampwidth == 2:
+                        frames = wf.readframes(num_frames)
+                        if not frames:
+                            print(f"  [Removing] Empty audio: {audio_path}")
+                            removed_count += 1
+                            continue
+
+                        samples = np.frombuffer(frames, dtype=np.int16)
+                        if samples.size == 0:
+                            print(f"  [Removing] Empty samples: {audio_path}")
+                            removed_count += 1
+                            continue
+
+                        normalized_samples = samples.astype(np.float32) / 32768.0
+                        max_abs_amplitude = float(
+                            np.max(np.abs(normalized_samples))
+                        )
+
+                        if max_abs_amplitude > max_amplitude_threshold:
+                            print(
+                                f"  [Removing] High amplitude: {audio_path} "
+                                f"(Peak: {max_abs_amplitude:.2f})"
+                            )
+                            removed_count += 1
+                            continue
+
+                        if max_abs_amplitude < min_amplitude_threshold:
+                            print(
+                                f"  [Removing] Low amplitude: {audio_path} "
+                                f"(Peak: {max_abs_amplitude:.2f})"
+                            )
+                            removed_count += 1
+                            continue
+                    else:
+                        print(
+                            f"  [Info] Skipping amplitude check for {audio_path} "
+                            f"(sample width: {sampwidth})"
+                        )
+
+            except wave.Error as e:
+                print(
+                    f"  [Warning] Could not read WAV after conversion/open: "
+                    f"{audio_path} ({e})"
+                )
+                removed_count += 1
+                continue
+
+            # Passed all checks
+            kept_lines.append(line + "\n")
+
+        except wave.Error as e:
+            print(
+                f"  [Warning] Could not read WAV file, skipping: {audio_path} ({e})"
+            )
+            removed_count += 1
+        except FileNotFoundError as e:
+            print(f"  [Warning] File disappeared, skipping: {audio_path} ({e})")
+            removed_count += 1
+        except Exception as e:
+            print(f"  [Error] An unexpected error occurred on line: {line}")
+            print(f"          {e}")
+            removed_count += 1
+
+    with open(output_path, "w", encoding="utf-8") as f:
+        f.writelines(kept_lines)
+
+    print("\n--- Processing Complete ---")
+    print(f"Total entries processed: {total_count}")
+    print(f"Entries kept: {len(kept_lines)}")
+    print(f"Entries removed (duration/amplitude/errors/dupes): {removed_count}")
+    print(f"Filtered list saved to: '{output_path}'")
+
+
+if __name__ == "__main__":
+    # Ensure numpy and ffmpeg are available.
+    filter_audio_list(
+        input_path="train_3_filtered.txt",
+        output_path="train_3_filtered_1.txt",
+        min_duration_sec=1.25,
+        max_duration_sec=10.0,
+        min_amplitude_threshold=0.02,
+        max_amplitude_threshold=0.93,
+        ffmpeg_path="ffmpeg",
+    )

Data/finalfinal.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8461bf253498a394a29ab7a017345f8d6ea09b57f7cd96e2705db5efd4740773
+size 47296629

Data/harry.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ebc9b1e699756f324fda10aad71573d4a4b026cf3fcaa29f42b55578b77f8937
+size 16283722

Data/luf_norm.py

@@ -0,0 +1,429 @@
+#!/usr/bin/env python3
+"""
+Batch LUFS normalization using ffmpeg (two-pass loudnorm), enforcing a peak
+limit (default linear 0.93) and converting to 44.1 kHz, mono, PCM 16-bit.
+Recursively processes all .wav files under a directory, atomically overwrites
+originals, and deletes any files that fail processing. Logs results to CSV.
+
+Important:
+ - Requires ffmpeg in PATH (ffmpeg loudnorm). Recommended ffmpeg >= 4.x.
+ - Requires Python packages: soundfile, numpy
+   Install with: pip install soundfile numpy
+
+Usage examples:
+  python luf_norm_ffmpeg_no_quarantine.py --input ../Dataset
+  python luf_norm_ffmpeg_no_quarantine.py --input ../Dataset --target -18.0 --peak 0.93
+  python luf_norm_ffmpeg_no_quarantine.py --input ../Dataset --workers 4
+"""
+from __future__ import annotations
+
+import argparse
+import csv
+import json
+import os
+import re
+import shutil
+import subprocess
+import tempfile
+import uuid
+from functools import partial
+from multiprocessing import Pool
+from pathlib import Path
+from typing import Dict, Optional
+
+import numpy as np
+import soundfile as sf
+
+# Regex used to extract JSON printed by ffmpeg loudnorm (it prints a JSON block to stderr)
+_JSON_RE = re.compile(r"\{[\s\S]*?\}", re.MULTILINE)
+
+
+def find_wav_files(root: Path):
+    return sorted(root.rglob("*.wav"))
+
+
+def _extract_loudnorm_json(stderr_text: str) -> Dict:
+    matches = list(_JSON_RE.finditer(stderr_text))
+    if not matches:
+        raise RuntimeError("ffmpeg loudnorm did not produce JSON measurement in stderr")
+    # Use last JSON block
+    js = matches[-1].group(0)
+    return json.loads(js)
+
+
+def ffmpeg_measure(in_path: Path, target_lufs: float, tp_db: float) -> Dict:
+    """
+    First-pass: run ffmpeg loudnorm with print_format=json to get measurement.
+    Returns parsed measurement dict.
+    """
+    cmd = [
+        "ffmpeg",
+        "-hide_banner",
+        "-nostats",
+        "-y",
+        "-i",
+        str(in_path),
+        "-af",
+        f"loudnorm=I={target_lufs}:TP={tp_db}:LRA=7:print_format=json",
+        "-ar",
+        "44100",
+        "-ac",
+        "1",
+        "-f",
+        "null",
+        "-",
+    ]
+    proc = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if proc.returncode != 0:
+        raise RuntimeError(f"ffmpeg measure failed (rc={proc.returncode}): {proc.stderr.splitlines()[-20:]}")
+    return _extract_loudnorm_json(proc.stderr)
+
+
+def ffmpeg_apply_second_pass(in_path: Path, out_path: Path, target_lufs: float, tp_db: float, measured: Dict, use_limiter: bool = False, limiter_db: Optional[float] = None):
+    """
+    Second pass: apply loudnorm with measured_* values to produce normalized output.
+    If use_limiter True, append an alimiter with limit=limiter_db (dB) after loudnorm.
+    Writes to out_path (wav).
+    """
+    # Required measured fields for the second pass
+    measured_I = measured.get("input_i")
+    measured_TP = measured.get("input_tp")
+    measured_LRA = measured.get("input_lra")
+    measured_thresh = measured.get("input_thresh")
+    target_offset = measured.get("target_offset", 0.0)
+
+    if None in (measured_I, measured_TP, measured_LRA, measured_thresh):
+        raise RuntimeError(f"Missing measured fields for second pass: {measured}")
+
+    # Build audio filter string
+    af = (
+        f"loudnorm=I={target_lufs}:TP={tp_db}:LRA=7:"
+        f"measured_I={measured_I}:measured_TP={measured_TP}:measured_LRA={measured_LRA}:"
+        f"measured_thresh={measured_thresh}:offset={target_offset}"
+    )
+    if use_limiter:
+        if limiter_db is None:
+            raise RuntimeError("Limiter requested but limiter_db is None")
+        # alimiter.limit expects a linear amplitude in [0.0625, 1], not dB.
+        linear_limit = float(10 ** (limiter_db / 20.0))
+        # Clamp to valid range to avoid ffmpeg errors
+        linear_limit = max(0.0625, min(1.0, linear_limit))
+        af = f"{af},alimiter=limit={linear_limit}"
+    cmd = [
+        "ffmpeg",
+        "-hide_banner",
+        "-nostats",
+        "-y",
+        "-i",
+        str(in_path),
+        "-af",
+        af,
+        "-ar",
+        "44100",
+        "-ac",
+        "1",
+        "-c:a",
+        "pcm_s16le",
+        str(out_path),
+    ]
+    proc = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    if proc.returncode != 0:
+        # include some stderr context for debugging
+        raise RuntimeError(f"ffmpeg apply failed (rc={proc.returncode}). stderr excerpt: {proc.stderr.splitlines()[-40:]}")
+
+
+def _unique_dest(dest: Path) -> Path:
+    """Return a unique path by appending a short uuid if dest exists."""
+    if not dest.exists():
+        return dest
+    stem = dest.stem
+    suffix = dest.suffix
+    parent = dest.parent
+    while True:
+        new_name = f"{stem}_{uuid.uuid4().hex[:8]}{suffix}"
+        candidate = parent / new_name
+        if not candidate.exists():
+            return candidate
+
+
+def atomic_replace(tmp_path: Path, dest: Path):
+    """
+    Atomically replace dest with tmp_path. tmp_path should exist.
+    """
+    try:
+        os.replace(str(tmp_path), str(dest))
+    except Exception:
+        # fallback to move
+        shutil.move(str(tmp_path), str(dest))
+
+
+def validate_audio_and_get_peak(path: Path) -> Dict:
+    """
+    Read audio file and return dict with sr, channels, peak, and whether samples are finite.
+    """
+    data, sr = sf.read(str(path), dtype="float32", always_2d=False)
+    channels = 1 if data.ndim == 1 else int(data.shape[1])
+    mono = data if data.ndim == 1 else np.mean(data, axis=1)
+    finite = bool(np.all(np.isfinite(mono)))
+    peak = float(np.max(np.abs(mono))) if mono.size else 0.0
+    return {"sr": int(sr), "channels": channels, "finite": finite, "peak": peak}
+
+
+def process_one(in_path: Path, target_lufs: float, peak_linear: float, workers_idx: int = 0) -> Dict:
+    """
+    Process a single file:
+      - 1st pass measure (ffmpeg loudnorm print json)
+      - 2nd pass apply loudnorm using measured values -> tmp out
+      - validate tmp out; if peak > peak_linear then re-run second pass with limiter
+      - if still invalid or non-finite, delete original and record error
+      - otherwise atomically overwrite original with tmp out
+    Returns a record dict for CSV logging.
+    """
+    rec = {
+        "path": str(in_path),
+        "input_i": None,
+        "input_tp": None,
+        "input_lra": None,
+        "input_thresh": None,
+        "target_offset": None,
+        "final_lufs": None,   # may be filled from measured output_i if present
+        "final_peak": None,
+        "sr": None,
+        "channels": None,
+        "deleted": False,
+        "error": None,
+    }
+
+    try:
+        # compute dBTP from linear peak
+        tp_db = 20.0 * np.log10(peak_linear) if peak_linear > 0 else -999.0
+        limiter_db = tp_db  # limiter in dB to use if reprocessing needed
+
+        # 1) Measure
+        measured = ffmpeg_measure(in_path, target_lufs, tp_db)
+
+        # record measured fields if available
+        rec["input_i"] = measured.get("input_i")
+        rec["input_tp"] = measured.get("input_tp")
+        rec["input_lra"] = measured.get("input_lra")
+        rec["input_thresh"] = measured.get("input_thresh")
+        rec["target_offset"] = measured.get("target_offset")
+
+        # Use a temp file in same directory to write output, then atomic replace
+        out_tmp = Path(in_path.parent) / (in_path.stem + f".tmp_{uuid.uuid4().hex[:8]}.wav")
+
+        # 2) Second pass: apply loudnorm (no limiter first)
+        try:
+            ffmpeg_apply_second_pass(in_path, out_tmp, target_lufs, tp_db, measured, use_limiter=False)
+        except Exception as e:
+            # second pass failed -> delete the original and record error
+            rec["error"] = f"ffmpeg second pass failed: {e}"
+            try:
+                in_path.unlink()
+                rec["deleted"] = True
+            except Exception as e2:
+                rec["error"] += f"; failed deleting original: {e2}"
+            # cleanup tmp if present
+            if out_tmp.exists():
+                try:
+                    out_tmp.unlink()
+                except Exception:
+                    pass
+            return rec
+
+        # 3) validate output
+        try:
+            info = validate_audio_and_get_peak(out_tmp)
+            rec["sr"] = info["sr"]
+            rec["channels"] = info["channels"]
+            rec["final_peak"] = float(info["peak"])
+            if not info["finite"]:
+                raise RuntimeError("Non-finite samples in ffmpeg output")
+        except Exception as e:
+            rec["error"] = (rec["error"] + "; " if rec["error"] else "") + f"Validation failed after second pass: {e}"
+            try:
+                in_path.unlink()
+                rec["deleted"] = True
+            except Exception as e2:
+                rec["error"] += f"; failed deleting original: {e2}"
+            if out_tmp.exists():
+                try:
+                    out_tmp.unlink()
+                except Exception:
+                    pass
+            return rec
+
+        # 4) If peak > peak_linear (allow tiny epsilon), re-run with limiter
+        eps = 1e-6
+        if rec["final_peak"] > peak_linear + eps:
+            # Run second pass again but append alimiter to ensure peaks are clamped
+            # write to a new temporary file
+            out_tmp2 = Path(in_path.parent) / (in_path.stem + f".tmp_lim_{uuid.uuid4().hex[:8]}.wav")
+            try:
+                ffmpeg_apply_second_pass(in_path, out_tmp2, target_lufs, tp_db, measured, use_limiter=True, limiter_db=limiter_db)
+            except Exception as e:
+                rec["error"] = (rec["error"] + "; " if rec["error"] else "") + f"ffmpeg limiter pass failed: {e}"
+                # delete original and cleanup
+                try:
+                    in_path.unlink()
+                    rec["deleted"] = True
+                except Exception as e2:
+                    rec["error"] += f"; failed deleting original: {e2}"
+                # cleanup temp files
+                for t in (out_tmp, out_tmp2):
+                    if t.exists():
+                        try:
+                            t.unlink()
+                        except Exception:
+                            pass
+                return rec
+
+            # validate out_tmp2
+            try:
+                info2 = validate_audio_and_get_peak(out_tmp2)
+                rec["sr"] = info2["sr"]
+                rec["channels"] = info2["channels"]
+                rec["final_peak"] = float(info2["peak"])
+                if not info2["finite"]:
+                    raise RuntimeError("Non-finite samples in limiter output")
+                # if still above limit, treat as error
+                if rec["final_peak"] > peak_linear + eps:
+                    raise RuntimeError(f"Peak still above limit after limiter: {rec['final_peak']:.6f} > {peak_linear:.6f}")
+            except Exception as e:
+                rec["error"] = (rec["error"] + "; " if rec["error"] else "") + f"Validation failed after limiter: {e}"
+                try:
+                    in_path.unlink()
+                    rec["deleted"] = True
+                except Exception as e2:
+                    rec["error"] += f"; failed deleting original: {e2}"
+                # cleanup temps
+                for t in (out_tmp, out_tmp2):
+                    if t.exists():
+                        try:
+                            t.unlink()
+                        except Exception:
+                            pass
+                return rec
+
+            # limiter output is good — replace the tmp pointer to limiter output for final replace
+            # remove the first tmp (if exists)
+            if out_tmp.exists():
+                try:
+                    out_tmp.unlink()
+                except Exception:
+                    pass
+            out_tmp = out_tmp2
+
+        # 5) All good: atomically overwrite original with out_tmp
+        try:
+            atomic_replace(out_tmp, in_path)
+        except Exception as e:
+            rec["error"] = (rec["error"] + "; " if rec["error"] else "") + f"Atomic replace failed: {e}"
+            # in case of a failure, try to delete original to avoid leaving inconsistent data (as requested)
+            try:
+                in_path.unlink()
+                rec["deleted"] = True
+            except Exception as e2:
+                rec["error"] += f"; failed deleting original after replace failure: {e2}"
+            if out_tmp.exists():
+                try:
+                    out_tmp.unlink()
+                except Exception:
+                    pass
+            return rec
+
+        # Optionally, ffmpeg's JSON sometimes includes output_i (final loudness).
+        # We didn't capture the apply-pass stderr here. If you want exact final LUFS, we could
+        # measure it using pyloudnorm or another ffmpeg measurement pass. For now, leave final_lufs empty.
+        rec["final_lufs"] = None
+
+        return rec
+
+    except Exception as e:
+        rec["error"] = str(e)
+        # On unexpected error, delete the original as user requested no quarantine
+        try:
+            in_path.unlink()
+            rec["deleted"] = True
+        except Exception as e2:
+            rec["error"] += f"; failed deleting original during error cleanup: {e2}"
+        return rec
+
+
+def write_csv(logfile: Path, records, fieldnames):
+    with open(logfile, "w", newline="", encoding="utf-8") as f:
+        writer = csv.DictWriter(f, fieldnames=fieldnames)
+        writer.writeheader()
+        for r in records:
+            writer.writerow({k: r.get(k) for k in fieldnames})
+
+
+def main():
+    p = argparse.ArgumentParser(description="Batch LUFS normalization using ffmpeg loudnorm (two-pass), delete bad files.")
+    p.add_argument("--input", "-i", default="../Dataset", help="Input root directory (recursive). Default ../Dataset")
+    p.add_argument("--target", "-t", type=float, default=-16.0, help="Target integrated LUFS (default -16.0)")
+    p.add_argument("--peak", type=float, default=0.93, help="Maximum linear peak allowed after processing (default 0.93)")
+    p.add_argument("--log", default="normalize_log.csv", help="CSV log filename")
+    p.add_argument("--workers", "-w", type=int, default=1, help="Number of parallel workers (default 1)")
+    args = p.parse_args()
+
+    input_root = Path(args.input).expanduser().resolve()
+    if not input_root.exists():
+        print(f"Input directory does not exist: {input_root}")
+        return
+
+    wav_files = find_wav_files(input_root)
+    if not wav_files:
+        print(f"No WAV files found under {input_root}")
+        return
+
+    print(f"Found {len(wav_files)} WAV files.")
+    print(f"Target LUFS = {args.target} dB. Peak limit (linear) = {args.peak} (TP ≈ {20.0 * np.log10(args.peak):.3f} dB).")
+    print("Converting to 44.1 kHz, mono, PCM_16. Bad files will be deleted (no quarantine).")
+    print(f"Workers: {args.workers}")
+
+    worker = partial(process_one, target_lufs=args.target, peak_linear=args.peak)
+
+    records = []
+    if args.workers == 1:
+        for pth in wav_files:
+            rec = worker(pth)
+            records.append(rec)
+            status = "OK" if not rec.get("error") else "ERR"
+            print(f"{pth}: {status} {'(deleted)' if rec.get('deleted') else ''} {rec.get('error') or ''}")
+    else:
+        with Pool(processes=args.workers) as pool:
+            for rec in pool.imap_unordered(worker, wav_files):
+                records.append(rec)
+                status = "OK" if not rec.get("error") else "ERR"
+                print(f"{rec.get('path')}: {status} {'(deleted)' if rec.get('deleted') else ''} {rec.get('error') or ''}")
+
+    # Write CSV
+    fieldnames = [
+        "path",
+        "input_i",
+        "input_tp",
+        "input_lra",
+        "input_thresh",
+        "target_offset",
+        "final_lufs",
+        "final_peak",
+        "sr",
+        "channels",
+        "deleted",
+        "error",
+    ]
+    write_csv(Path(args.log), records, fieldnames)
+
+    total = len(records)
+    errors = sum(1 for r in records if r.get("error"))
+    deleted = sum(1 for r in records if r.get("deleted"))
+    print("\nSummary:")
+    print(f"Total files processed: {total}")
+    print(f"Files with error: {errors}")
+    print(f"Files deleted: {deleted}")
+    print(f"CSV log written to: {args.log}")
+
+
+if __name__ == "__main__":
+    main()

Data/merged_train_test.txt

@@ -0,0 +1,23 @@
+Dataset/wavs/czytadlo_ferenc/9788367950787_005_Konwicki_Czytadlo_segment_218.wav|patʃˈɨwa na krˈavɛɲdʑ zbˈɔtʃa ujˈɛntɔ̃ w ɡˈɛ̃stfɛ pˈustɨx kʃˈakuf.|0|Patrzyła na krawędź zbocza ujętą w gęstwę pustych krzaków.
+Dataset/wavs/lew_ferenc/12_segment_391.wav|tɔ znˈatʃɲʲɛ pʃɨspʲˈɛʃa pˌɔtpisˈaɲʲɛ uɡˈɔdɨ mʲˌɛndʑɨ vwˌaɕtɕitɕˈɛlɛm zˌapadˌajɔntsˈɛɡɔ ɕɛ dˈɔmu a kɔpˈalɲʲɔ̃.|0|To znacznie przyspiesza podpisanie ugody między właścicielem zapadającego się domu a kopalnią.
+Dataset/wavs/ryz_ferenc/9788367950190_15_Hlasko_Palcie-ryz-kazdego-dnia_segment_149.wav|pʃˌɨpɔmˈinaw mu ɔ tʃˈɨmɕ, tsˈɔ ɕɛ ɲʲɛ stˈawɔ.|0|Przypominał mu o czymś, co się nie stało.
+Dataset/wavs/czytadlo_ferenc/9788367950787_006_Konwicki_Czytadlo_segment_153.wav|pɔtʃˈɛdwɛm dɔ naʃˈɛɡɔ bufˈɛtu dˈɔɕtɕ sfˌatɨɡˌɔvanˈɛɡɔ pʃˈɛstɛ fʃˈɨstkʲɛ lˈata.|0|Podszedłem do naszego bufetu dość sfatygowanego przez te wszystkie lata.
+Dataset/wavs/czytadlo_ferenc/9788367950787_004_Konwicki_Czytadlo_segment_105.wav|ʒˌɛtʃɨvˈiɕtɕɛ mˈiwɔ. lˈubjɛ tɛn vˈus.|0|Rzeczywiście miło. Lubię ten wóz.
+Dataset/wavs/czytadlo_ferenc/9788367950787_007_Konwicki_Czytadlo_segment_197.wav|a jɛj sˈɛn tɔ zɡˈɔda i pʃˌɨzvɔlˈɛɲʲɛ.|0|a jej sen to zgoda i przyzwolenie.
+Dataset/wavs/czytadlo_ferenc/9788367950787_008_Konwicki_Czytadlo_segment_401.wav|ʒˈɛ jˈa, lˈɛkkɔː ˌarxaˈitʃnɨ i mˈɔtsnɔ naˈivnɨ,|0|Że ja, lekko archaiczny i mocno naiwny,
+Dataset/wavs/lew_ferenc/05_segment_296.wav|kʲˈɛdɨɕ na bˈɔbrɛk jɛxˈawɔ ɕɛ pɔ lˈɛpʃɛ ʒˈɨtɕɛ.|0|Kiedyś na Bobrek jechało się po lepsze życie.
+Dataset/wavs/urwisko_ferenc/22_segment_14.wav|ɡˈɛstɛm zaprˈɔɕiw ɡˈɔ dɔ vˈɔzu. bɔrˈɛvitʃ ˈuɕad na fɔtˈɛlu pˌasaʒˈɛra i zˈamknɔw dʒvˈi, ktˌurɛ pʃɨlɡnˈɛwɨ ɡwˈuxɔ dɔː uʃtʃˈɛlki.|0|Gestem zaprosił go do wozu. Borewicz usiadł na fotelu pasażera i zamknął drzwi, które przylgnęły głucho do uszczelki.
+Dataset/wavs/czytadlo_ferenc/9788367950787_012_Konwicki_Czytadlo_segment_97.wav|zatʃˈɨnam ˌɔdʑɨskˈivatɕ pˈamjɛɲtɕ. pʃɨpˈɔmnɛ sˈɔbjɛ fʃˈɨstkɔ.|0|Zaczynam odzyskiwać pamięć. Przypomnę sobie wszystko.
+Dataset/wavs/czytadlo_ferenc/9788367950787_001_Konwicki_Czytadlo_segment_593.wav|pˈadaw dˈɛʃtʃ.|0|Padał deszcz.
+Dataset/wavs/czytadlo_ferenc/9788367950787_009_Konwicki_Czytadlo_segment_206.wav|pɔdɲʲˈɔswɛm ɕɛ tɕˈɛ̃ʃkɔ z wˈuʃka.|0|Podniosłem się ciężko z łóżka.
+Dataset/wavs/ryz_ferenc/9788367950190_16_Hlasko_Palcie-ryz-kazdego-dnia_segment_131.wav|tʃɨ ɲʲɛ jɛst tˈak? zapˈɨtaw.|0|Czy nie jest tak? zapytał.
+Dataset/wavs/lew_ferenc/07_segment_247.wav|pɔwˈɔvɛ mʲɛʃkˈaɲtsuf stanˈɔvi lˈudnɔɕtɕ nˌapwɨvˈɔva, tɔ jɛst rˌɛpatrjˈaɲtɕi i ˌɔɕɛdlˈɛɲtsɨ.|0|Połowę mieszkańców stanowi ludność napływowa, to jest repatrianci i osiedleńcy.
+Dataset/wavs/czytadlo_ferenc/9788367950787_003_Konwicki_Czytadlo_segment_259.wav|w nˈaɡwɛj pʃˌɨtɔmnˈɔɕtɕi pˌɔdʑivʲˈawɛm fspaɲʲˈawɛ prɔpˈɔrtsjɛ i zdˌumjɛvajˈɔntsɔ̃ ɡˌɛɔmɛtrˈɨtʃnɔɕtɕ pʲˈɛrɕi ˌaɲi za dˈuʒɨx, ˌaɲi za mˈawɨx.|0|W nagłej przytomności podziwiałem wspaniałe proporcje i zdumiewającą geometryczność piersi ani za dużych, ani za małych.
+Dataset/wavs/lew_ferenc/05_segment_360.wav|ˌɔtʃɨvˈiɕtɕɛ, mˈɨɕmɨ ix ʃkɔlˈili, bˈɔ nˈatɔ dɔstˈaliɕmˌɨ pʲɛɲʲˈɔndʑɛ.|0|Oczywiście, myśmy ich szkolili, bo na to dostaliśmy pieniądze.
+Dataset/wavs/urwisko_ferenc/15_segment_100.wav|miltʃˈawa. dˈam tɕi nˈamjar na jˈɛɡɔ bˈɨwɔ̃ ʒˈɔnɛ.|0|Milczała. Dam ci namiar na jego byłą żonę.
+Dataset/wavs/urwisko_ferenc/49_segment_65.wav|pˌɔbwɔɲdʑˈiwa, ˌalɛ tɔ ɕɛ zdaʒˈawɔ najlˈɛpʃɨm.|0|Pobłądziła, ale to się zdarzało najlepszym.
+Dataset/wavs/czytadlo_ferenc/9788367950787_008_Konwicki_Czytadlo_segment_293.wav|ɔpˈɔvʲɛts mi tsˈɔɕ ɔ sˈɔbjɛ, ʃɛpnˈɛwa.|0|Opowiedz mi coś o sobie, szepnęła.
+Dataset/wavs/czytadlo_ferenc/9788367950787_007_Konwicki_Czytadlo_segment_118.wav|naxˈɔdʑi mɲʲɛ pɔlˈitsja.|0|Nachodzi mnie policja.
+Dataset/wavs/urwisko_ferenc/63_segment_128.wav|tʃɨ ɕɲˈi mu ɕɛ zuzˈanna?|0|Czy śni mu się Zuzanna?
+Dataset/wavs/czytadlo_ferenc/9788367950787_005_Konwicki_Czytadlo_segment_502.wav|a mɨ ftˈɛdɨ vidʑˈimɨ pʃɛkwˈutɔ̃ tɕˈɛmnɔɕtɕ.|0|A my wtedy widzimy przekłutą ciemność.
+Dataset/wavs/urwisko_ferenc/43_segment_78.wav|ˈɔlɡʲɛrd vˈɨjɔw tɛlˈɛfɔn i vˈɨbraw pˌɔwɔntʃˈɛɲʲɛ. pʃɨwˈɔʒɨw apˈarad dɔː ˈuxa.|0|Olgierd wyjął telefon i wybrał połączenie. Przyłożył aparat do ucha.

Data/merged_val.txt

@@ -0,0 +1,106 @@
+Dataset/wavs/urwisko_ferenc/58_segment_27.wav|ɲʲɛ bˈardʑɔ vʲˈɛdʑaw, tsˈɔ ɕɛ dʑˈɛjɛ nˈatɨm parkˈiŋɡu i tʃɨ pɔvˈiɲʲɛn tsˈɔɕ s tɨm zrˈɔbʲitɕ.|0|Nie bardzo wiedział, co się dzieje na tym parkingu i czy powinien coś z tym zrobić.
+Dataset/wavs/ryz_ferenc/9788367950190_09_Hlasko_Palcie-ryz-kazdego-dnia_segment_293.wav|ˈandɛrsɔn i sˈusan patʃˈɨli za ɲˈimi.|0|Anderson i Susan patrzyli za nimi.
+Dataset/wavs/ryz_ferenc/9788367950190_06_Hlasko_Palcie-ryz-kazdego-dnia_segment_50.wav|ʑlˈɛ. mˈɔi kljˈɛɲtɕi tɔ vwˈaɕɲʲɛ tˈatsɨ jˈak tˈɨ.|0|Źle. Moi klienci to właśnie tacy jak ty.
+Dataset/wavs/urwisko_ferenc/65_segment_129.wav|ˌalɛ najɡˈɔrʃɛ bˌɨwɨ ɔstˈatɲʲɛ swˈɔva rˌɛpɔrtˈaʒu.|0|Ale najgorsze były ostatnie słowa reportażu.
+Dataset/wavs/czytadlo_ferenc/9788367950787_010_Konwicki_Czytadlo_segment_107.wav|pɔʃˈɛdwɛm f strˈɔnɛ skˈarpɨ.|0|Poszedłem w stronę skarpy.
+Dataset/wavs/lew_ferenc/09_segment_238.wav|ɲʲˌɛvɨbrˈanɨ vˈɛŋɡʲɛl, ktˌurɨ mˈɔʒna bˌɨwɔ vˈɨbratɕ, tɔː ɔɡrˈɔmna strˈata dla pʃɛmˈɨswu, dla ˌɔbɨvatˈɛli.|0|Niewybrany węgiel, który można było wybrać, to ogromna strata dla przemysłu, dla obywateli.
+Dataset/wavs/urwisko_ferenc/23_segment_34.wav|pˈaxɲʲaw dˈɔbrɔ̃ vˈɔdɔ̃ kɔlˈɔɲskɔ̃.|0|Pachniał dobrą wodą kolońską.
+Dataset/wavs/czytadlo_ferenc/9788367950787_010_Konwicki_Czytadlo_segment_422.wav|a dˈɔkɔnt jɛdʑˈɛmɨ? a, zɔbˈatʃɨ pˈan na mʲˈɛjstsu.|0|A dokąd jedziemy? A, zobaczy pan na miejscu.
+Dataset/wavs/czytadlo_ferenc/9788367950787_009_Konwicki_Czytadlo_segment_208.wav|xˈalɔ, pˌɔvʲɛdʑˈawɛm sˈɛnnɨm ɡwˈɔsɛm.|0|Halo, powiedziałem sennym głosem.
+Dataset/wavs/lew_ferenc/11_segment_292.wav|uwˈamɛk sɛkˈundɨ.|0|Ułamek sekundy.
+Dataset/wavs/czytadlo_ferenc/9788367950787_006_Konwicki_Czytadlo_segment_293.wav|xˈɔtɕ ˈinɲi zmʲɛɲʲˈali ɕɛ tsˈɔ pˈuw kˌilɔmˈɛtra.|0|Choć inni zmieniali się co pół kilometra.
+Dataset/wavs/urwisko_ferenc/21_segment_45.wav|tˈɨlɛ tˈɨlkɔ, ʒˈɛ xwˈɔp ɲʲˈɛma mɔtˈɨvu. jˈɛɕli dˈɔbʒɛ zrˌɔzumjˈawɛm, rɔzvˈɔdʑiw ɕɛ z ʒˈɔnɔ̃, ktˌura mʲˈawa tfˈardɛ dɔvˈɔdɨ zdrˈadɨ.|0|Tyle tylko, że chłop nie ma motywu. Jeśli dobrze zrozumiałem, rozwodził się z żoną, która miała twarde dowody zdrady.
+Dataset/wavs/ryz_ferenc/9788367950190_07_Hlasko_Palcie-ryz-kazdego-dnia_segment_199.wav|tʃɨ ɲʲɛ bˌɨwɔ tˈak?|0|Czy nie było tak?
+Dataset/wavs/ryz_ferenc/9788367950190_13_Hlasko_Palcie-ryz-kazdego-dnia_segment_46.wav|ʃˈɛfjɛ, ˌɔɲi ɕɛ ˌɔbraʒˈajɔ̃, kʲˈɛdɨ ix ɕɛ bjˈɛʒɛ jˈɛdnɨɣ za drˈuɡʲix. tˈak? ˌɔbraʒˈajɔ̃ ɕɛ?|0|Szefie, oni się obrażają, kiedy ich się bierze jednych za drugich. Tak? Obrażają się?
+Dataset/wavs/lew_ferenc/04_segment_213.wav|tsudˈɔvnɛ zdaʒˈɛɲʲa tɔ mˈɔɡwɨbˌɨ mʲˈɛtɕ mʲˈɛjstsɛ w bˈajtsɛ ɔ ɕlˈɔ̃skim kɔptɕˈuʃku.|0|Cudowne zdarzenia to mogłyby mieć miejsce w bajce o śląskim kopciuszku.
+Dataset/wavs/lew_ferenc/09_segment_391.wav|pɔt stʃˈɛxɛ jˈɛdnɛj z vʲˈɛjskix xˈad zˌaxɔdʑˈiwɨ vˈɨʃʃɛ sfˈɛrɨ pɔ ɕfʲɛʒˈutkɔ̃ ɕmʲɛtˈaŋkɛ, pɔ mlˈɛkɔ.|0|Pod strzechę jednej z wiejskich chat zachodziły wyższe sfery po świeżutką śmietankę, po mleko.
+Dataset/wavs/urwisko_ferenc/28_segment_42.wav|ilˈɛkrɔtɕ spˌɔtɨkˈawam ɕɛ s kˈaɕkɔ̃, vidʑˈawam, ʒˈɛ ɲʲɛ ukwˈada jɛj ɕɛ w mawʒˈɛɲstfʲɛ.|0|Ilekroć spotykałam się z Kaśką, widziałam, że nie układa jej się w małżeństwie.
+Dataset/wavs/czytadlo_ferenc/9788367950787_009_Konwicki_Czytadlo_segment_159.wav|tɛn vʲˈɛlki vˈir mɲʲɛ vˈɛssaw.|0|Ten wielki wir mnie wessał.
+Dataset/wavs/lew_ferenc/13_segment_137.wav|pˌɔtstavˈili ˈawta, vɨɡnˈali lˈudʑi s xawˈupɨ, wˌadɔvˈali i vˌɨvɔzˈili na xɔtˈɛlɛ, jˈak ktˈɔɕ ɲʲˈɛmʲaw rɔdʑˈinɨ w ˌɔkɔlˈitsɨ.|0|Podstawili auta, wygnali ludzi z chałupy, ładowali i wywozili na hotele, jak ktoś nie miał rodziny w okolicy.
+Dataset/wavs/urwisko_ferenc/32_segment_27.wav|ftˈɛdɨ stˈara nˈɔkja xˈɛrman...|0|Wtedy stara Nokia Herman...
+Dataset/wavs/ryz_ferenc/9788367950190_13_Hlasko_Palcie-ryz-kazdego-dnia_segment_206.wav|ɲʲˈɛpij za dˈuʒɔ.|0|Nie pij za dużo.
+Dataset/wavs/czytadlo_ferenc/9788367950787_009_Konwicki_Czytadlo_segment_544.wav|ja bˌɛndɛ tfˈɔim ʒˈɨtɕɛm, pˌɔvʲɛdʑˈawa pʃɛt ɕˈɛbjɛ i tfˈɔjɔ̃ ɕmʲˈɛrtɕɔ̃.|0|Ja będę Twoim życiem, powiedziała przed siebie i Twoją śmiercią.
+Dataset/wavs/ryz_ferenc/9788367950190_01_Hlasko_Palcie-ryz-kazdego-dnia_segment_179.wav|tˈak tʃɨ ˈɔvak papʲˈɛrɨ sɔ̃ sfˌawʃɔvˈanɛ, a tˈu jɛzd vˈɔlnɨ krˈaj.|0|Tak czy owak papiery są sfałszowane, a tu jest wolny kraj.
+Dataset/wavs/ryz_ferenc/9788367950190_07_Hlasko_Palcie-ryz-kazdego-dnia_segment_194.wav|tˌɨlkɔ ta dʑɛftʃˈɨna ɲʲɛ pˈɔʃwa za tˈɔbɔ̃, tˈak jˈak tɨ sˈɔbjɛ tɔ vˌɨɔbraʒˈawɛɕ.|0|Tylko ta dziewczyna nie poszła za tobą, tak jak ty sobie to wyobrażałeś.
+Dataset/wavs/ryz_ferenc/9788367950190_10_Hlasko_Palcie-ryz-kazdego-dnia_segment_28.wav|w jˈakimɕ sˈɛnɕɛ.|0|W jakimś sensie.
+Dataset/wavs/czytadlo_ferenc/9788367950787_006_Konwicki_Czytadlo_segment_81.wav|ˌalɛ pʃˌɛtɕɛʃ miljˈardɨ lˈudʑi ɔ tsˈɔɕ ɕɛ mˈɔdlɔ̃ f tɛj xfˈili. ktˈɔ mˈa spˈɛwɲitɕ tɛ nˈaʃɛ ʒɨtʃˈɛɲʲa?|0|Ale przecież miliardy ludzi o coś się modlą w tej chwili. Kto ma spełnić te nasze życzenia?
+Dataset/wavs/lew_ferenc/12_segment_275.wav|xtɕˈawbɨm vʲˈɛdʑɛtɕ, tʃɨ pɔt nˈaʃɔ̃ ulˈitsɔ̃ ftɕˈɔ̃ʃ kˈɔpjɔ̃ vˈɛŋɡʲɛl.|0|Chciałbym wiedzieć, czy pod naszą ulicą wciąż kopią węgiel.
+Dataset/wavs/ryz_ferenc/9788367950190_10_Hlasko_Palcie-ryz-kazdego-dnia_segment_202.wav|ɲʲˈɛbuj ɕɛ, zɔstˈavjɛ tɕi jˈɔ̃.|0|Nie bój się, zostawię ci ją.
+Dataset/wavs/ryz_ferenc/9788367950190_09_Hlasko_Palcie-ryz-kazdego-dnia_segment_68.wav|tɔ zˈafʃɛ dˈɔbʒɛ zrˈɔbʲitɕ pɔvʲˈɛdʑaw ˈandɛrsɔn. pɔwˈɔʒɨw ɕɛ na plˈɛtsax i zawˈɔʒɨw ramjˈɔna pɔd ɡwˈɔvɔ̃.|0|– To zawsze dobrze zrobić – powiedział Anderson. Położył się na plecach i założył ramiona pod głową.
+Dataset/wavs/czytadlo_ferenc/9788367950787_009_Konwicki_Czytadlo_segment_542.wav|tʃˈɛba ɕɛ tɕˈɛʃɨtɕ kˈaʒdɨm dɲʲˈɛm. ɲʲɛ vʲˈɛmɨ, ˈilɛ nˈam iɣ zɔstˈawɔ.|0|Trzeba się cieszyć każdym dniem. Nie wiemy, ile nam ich zostało.
+Dataset/wavs/lew_ferenc/05_segment_212.wav|na ulˈitsɛ vˌɨjɛʒdʒˈajɔ̃ nˈɔvɛ sˌamɔxˈɔdɨ, ktˌurɛ za xfˈilɛ stˈanɔ̃ w ɡarˈaʒax s pustˈɨmi bakˈami.|0|Na ulicę wyjeżdżają nowe samochody, które za chwilę staną w garażach z pustymi bakami.
+Dataset/wavs/czytadlo_ferenc/9788367950787_007_Konwicki_Czytadlo_segment_268.wav|jˈak trafjˈawam i dlatʃˈɛɡɔ f tɔ nˈɔvɛ mʲˈɛjstsɛ?|0|Jak trafiałam i dlaczego w to nowe miejsce?
+Dataset/wavs/urwisko_ferenc/29_segment_4.wav|ɕˈɛŋɡnɔw pɔ kɔmˈurkɛ.|0|Sięgnął po komórkę.
+Dataset/wavs/urwisko_ferenc/43_segment_8.wav|xˈɛrman ɕɛ ˌutɕɛʃˈɨwa, ʒˈɛ trafˈili na ɡadˈuwɛ, ˌalɛ z dˌɔɕfʲattʃˈɛɲʲa vʲɛdʑˈawa, ʒˈɛ pɔvˈinna pʃˈɛjɔɲtɕ kɔntrˈɔlɛ nat jˈɛɡɔ swˌɔvɔtˈɔkʲɛm.|0|Herman się ucieszyła, że trafili na gadułę, ale z doświadczenia wiedziała, że powinna przejąć kontrolę nad jego słowotokiem.
+Dataset/wavs/czytadlo_ferenc/9788367950787_001_Konwicki_Czytadlo_segment_173.wav|a ja tˈak strˈaʃɲʲɛ ɲʲɛ lˈubjɛ rˌɔspɔtʃɨnˈaɲʲa mˈɨɕli ɔt tˌɛɡɔ nˈɔ tsˈuʃ.|0|A ja tak strasznie nie lubię rozpoczynania myśli od tego no cóż.
+Dataset/wavs/ryz_ferenc/9788367950190_05_Hlasko_Palcie-ryz-kazdego-dnia_segment_361.wav|tʃwˈɔvʲɛk, kturˌɛɡɔ mi brˌakɔvˈawɔ dɔ kɔmplˈɛtu.|0|Człowiek, którego mi brakowało do kompletu.
+Dataset/wavs/czytadlo_ferenc/9788367950787_008_Konwicki_Czytadlo_segment_239.wav|pˌɔwɔʒˈɨwɛm ɕɛ, ɔkrˈɨwɛm jˈɔ̃ ramjˈɛɲʲɛm.|0|Położyłem się, okryłem ją ramieniem.
+Dataset/wavs/ryz_ferenc/9788367950190_14_Hlasko_Palcie-ryz-kazdego-dnia_segment_106.wav|ɔ ɕmʲˈɛrtɕi tfɔjˈɛɡɔ brˈata vʲˈɛʃ tˌɨlkɔ dlatˈɛɡɔ, ʒˈɛ zɡʲˈinɔw f sˌamɔlˈɔtɕɛ i ɲʲɛ mˈɔʒɛʃ mu pʃɛbˈatʃɨtɕ, ʒˈɛ rɔzvˈaliw dˈɔbrɨ samˈɔlɔt.|0|O śmierci twojego brata wiesz tylko dlatego, że zginął w samolocie i nie możesz mu przebaczyć, że rozwalił dobry samolot.
+Dataset/wavs/czytadlo_ferenc/9788367950787_011_Konwicki_Czytadlo_segment_291.wav|zˌabwɔŋkˈawa ɕɛ dɔ tɛj ɡalˈaktɨkˌi i ɲʲɛ mˈɔʒɛ ɕɛ z ɲʲˈɛj vˈɨrvatɕ.|0|Zabłąkała się do tej galaktyki i nie może się z niej wyrwać.
+Dataset/wavs/ryz_ferenc/9788367950190_16_Hlasko_Palcie-ryz-kazdego-dnia_segment_145.wav|a jˈɛɕli ja tɕi ɲʲɛ pɔzvˈɔlɛ?|0|A jeśli ja ci nie pozwolę?
+Dataset/wavs/czytadlo_ferenc/9788367950787_004_Konwicki_Czytadlo_segment_59.wav|bˈɔ pʃˌɨpɔmɲʲˈawɛm sˈɔbjɛ sfˈuj dˈɔm, ktˌurɨ bˈɨw vʲˈɛlkɔ̃ kˌamjɛɲˈitsɔ̃ jˈak mˈɔnd blˈants.|0|Bo przypomniałem sobie swój dom, który był wielką kamienicą jak Mont Blanc.
+Dataset/wavs/ryz_ferenc/9788367950190_12_Hlasko_Palcie-ryz-kazdego-dnia_segment_56.wav|i ʒˈɛ ɲʲɛ stˈartʃɨ tɕi ʒˈɨtɕa, ˈabɨ ɕɛ ɲʲˈɔ̃ natɕˈɛʃɨtɕ.|0|I że nie starczy ci życia, aby się nią nacieszyć.
+Dataset/wavs/lew_ferenc/11_segment_270.wav|pɔvˈɔli zɲikˈamɨ s pɔvʲˈɛʃxɲi ʑˈɛmi. zɔstˈajɔ̃ zˌarɔɕɲʲˈɛntɛ kʃɛvˈami bɔtʃɲˈitsɛ.|0|Powoli znikamy z powierzchni ziemi. Zostają zarośnięte krzewami bocznice.
+Dataset/wavs/czytadlo_ferenc/9788367950787_005_Konwicki_Czytadlo_segment_207.wav|jˈakaɕ pʃˈɨkrɔʑdʑ zvʲɔ̃zˈana s tˈamtɔ̃ nˈɔtsɔ̃.|0|Jakaś przykrość związana z tamtą nocą.
+Dataset/wavs/ryz_ferenc/9788367950190_03_Hlasko_Palcie-ryz-kazdego-dnia_segment_297.wav|f pɔʒˈɔntku, pɔvʲˈɛdʑaw ɡˈɛrs fˈirst lˈɔvɛ.|0|W porządku, powiedział Gers First Love.
+Dataset/wavs/czytadlo_ferenc/9788367950787_001_Konwicki_Czytadlo_segment_500.wav|tsˈɔ mɲʲɛ ˌɔddʑɛlˈiwɔː ɔt mˈɛɡɔ ɡˌasnɔntsˈɛɡɔ ʒɨvˈɔta?|0|Co mnie oddzieliło od mego gasnącego żywota?
+Dataset/wavs/czytadlo_ferenc/9788367950787_007_Konwicki_Czytadlo_segment_345.wav|xˈɔdʑ znˈuf ɲʲɛ tˈaka mwˈɔda, jˈaɡbɨ ta mˈɔja kɔbjˈɛta tʃɨmˈawa tsˈɔɕ swɔtkʲˈɛɡɔ na jɛ̃zˈɨku ˈalbɔ pˌɔdɲʲɛbjˈɛɲʲu.|0|Choć znów nie taka młoda, jakby ta moja kobieta trzymała coś słodkiego na języku albo podniebieniu.
+Dataset/wavs/lew_ferenc/13_segment_215.wav|mˈɔʒna bˈɨ mu nˈadatɕ pˌɔvaʒɲʲˈɛjʃɔ̃ rˈaŋɡɛ, fprɔvˈadʑidʑ dɔ ɕtɕiswˈɛɡɔ tsˈɛntrum.|0|Można by mu nadać poważniejszą rangę, wprowadzić do ścisłego centrum.
+Dataset/wavs/urwisko_ferenc/36_segment_107.wav|pʃˈɛʃli pʃˈɛs pˈɔkuj na tˈaras. rɔzɡʒˈanɛ pwˈɨtki zɡʒɨtˈawɨ pɔt ix stɔpˈami.|0|Przeszli przez pokój na taras. Rozgrzane płytki zgrzytały pod ich stopami.
+Dataset/wavs/czytadlo_ferenc/9788367950787_004_Konwicki_Czytadlo_segment_197.wav|uɕmʲˈɛxatɕ ɕɛ dɔ mˌitskʲɛvˈitʃa.|0|Uśmiechać się do Mickiewicza.
+Dataset/wavs/czytadlo_ferenc/9788367950787_001_Konwicki_Czytadlo_segment_476.wav|ja ɲʲɛ jˌɛstɛm ˌɛrɔtɔmˈanɛm. ɲʲɛ lˈɛtsɛ jˈak pʲˈɛz za kˈaʒdɔ̃ sˈukɔ̃.|0|Ja nie jestem erotomanem. Nie lecę jak pies za każdą suką.
+Dataset/wavs/urwisko_ferenc/51_segment_17.wav|ftʃˈɛɕɲʲɛj pʃˌɨɡlɔndˈawa ɕɛ mˈapʲɛ ɡɔˈɔɡlɛ i zdjˈɛɲtɕɔm sˌatɛlitˈarnɨm kaʃtʃˈɔrka.|0|Wcześniej przyglądała się mapie Google i zdjęciom satelitarnym kaszczorka.
+Dataset/wavs/ryz_ferenc/9788367950190_15_Hlasko_Palcie-ryz-kazdego-dnia_segment_75.wav|— jˌɛstɛɕ jˈasna, a mɲʲɛ zˈafʃɛ muvʲˈɔnɔ, ʒˈɛ irlˈantki sɔ̃ tɕˌɛmnɔvwˈɔsɨ.|0|— Jesteś jasna, a mnie zawsze mówiono, że Irlandki są ciemnowłosy.
+Dataset/wavs/ryz_ferenc/9788367950190_02_Hlasko_Palcie-ryz-kazdego-dnia_segment_392.wav|— pˌɔvʲɛdʑˈawɛm jˈuʃ, pˈanu, pˈaɲi i ja jɛstˌɛɕmɨ tɨm sˈamɨm tʃwɔvʲˈɛkʲɛm.|0|— Powiedziałem już, panu, pani i ja jesteśmy tym samym człowiekiem.
+Dataset/wavs/ryz_ferenc/9788367950190_07_Hlasko_Palcie-ryz-kazdego-dnia_segment_39.wav|ˌalɛ pʃˈɛttɛm xtɕˈawbɨm fstˈɔmpidʑ dɔ sklˈɛpu z brˈɔɲʲɔ̃. tɔ pɔ drˈɔdʑɛ. xtsˈɛ kˈupitɕ karˈabʲin.|0|Ale przedtem chciałbym wstąpić do sklepu z bronią. To po drodze. Chcę kupić karabin.
+Dataset/wavs/ryz_ferenc/9788367950190_02_Hlasko_Palcie-ryz-kazdego-dnia_segment_9.wav|i pɔwˈɔʒɨw ɕɛ, pˌɔdwɔʒˈɨfʃɨ vʲatrˈufkɛ pɔd ɡwˈɔvɛ i ˌɔdvrutɕˈifʃɨ tfˈaʒ dɔ ɕtɕˈanɨ.|0|I położył się, podłożywszy wiatrówkę pod głowę i odwróciwszy twarz do ściany.
+Dataset/wavs/czytadlo_ferenc/9788367950787_008_Konwicki_Czytadlo_segment_431.wav|ɔbjˈɔwɛm jˈɔ̃ za ʃˈɨjɛ tˈak ˌɛnɛrɡʲˈitʃɲʲɛ, ʒˈɛ ˈaʒ zˌawaskɔtˈawa mˈuj bˈɔk kˌɔlɛbjɔntsˈɨmi ɕɛ sutkˈami. — tʃɨ tɨ ɕɛ mɲʲɛ ɲʲɛ bˈɔiʃ? — ʃɛpnˈɔwɛm.|0|Objąłem ją za szyję tak energicznie, że aż załaskotała mój bok kolebiącymi się sutkami. — Czy ty się mnie nie boisz? — szepnąłem.
+Dataset/wavs/urwisko_ferenc/38_segment_31.wav|zˌavaxˈawa ɕɛ, lˈɛtʃ ˌɔstatˈɛtʃɲʲɛ uɕˈadwa ˈɔbɔk.|0|Zawahała się, lecz ostatecznie usiadła obok.
+Dataset/wavs/urwisko_ferenc/24_segment_15.wav|pˈan sprˈavdʑi i zadʑvˈɔɲi.|0|Pan sprawdzi i zadzwoni.
+Dataset/wavs/ryz_ferenc/9788367950190_10_Hlasko_Palcie-ryz-kazdego-dnia_segment_66.wav|tɨ ˈɔtɨm vʲˈɛʃ rˈuvɲʲɛ dˈɔbʒɛ jˈak i jˈa. lˈataʃ i dwˈuʒɛj i lˈɛpʲɛj ˈɔdɛ mɲʲɛ. pʃˈɨkrɔ mi, ʒˈɛ mˈuʃɛ pɔvʲˈɛdʑɛtɕ tɕi kɔmplˈɛmɛnt.|0|Ty o tym wiesz równie dobrze jak i ja. Latasz i dłużej i lepiej ode mnie. Przykro mi, że muszę powiedzieć ci komplement.
+Dataset/wavs/urwisko_ferenc/49_segment_78.wav|pˈɔza tɨm jˌɛstɛm xɔlˈɛrɲʲɛ tɕˈɛkaf, tsˈɔ kˌɔmbʲinˈujɛʃ.|0|Poza tym jestem cholernie ciekaw, co kombinujesz.
+Dataset/wavs/urwisko_ferenc/59_segment_77.wav|pˌɔpatʃˈɨwa na bjɛɡwˈɛɡɔː i tʃɛkˈawa.|0|Popatrzyła na biegłego i czekała.
+Dataset/wavs/ryz_ferenc/9788367950190_15_Hlasko_Palcie-ryz-kazdego-dnia_segment_81.wav|tʃɨ bɛɲdʑˌɛmɨ tˈɛras mˈuvitɕ ɔ sˈukɲʲax?|0|Czy będziemy teraz mówić o sukniach?
+Dataset/wavs/ryz_ferenc/9788367950190_03_Hlasko_Palcie-ryz-kazdego-dnia_segment_97.wav|ɔtkwˈadaj sfˈɔjɛ pajˈɔŋki, a ja bˌɛndɛ ɔtkwˈadaw sfˈɔjɛ ɕmʲˈɛtɕɛ w lˈɛvɔ.|0|Odkładaj swoje pająki, a ja będę odkładał swoje śmiecie w lewo.
+Dataset/wavs/czytadlo_ferenc/9788367950787_003_Konwicki_Czytadlo_segment_440.wav|tʃˈas raptˈɔvnɨx navrˈutsɛɲ, pˌɔmɨɕlˈawɛm.|0|Czas raptownych nawróceń, pomyślałem.
+Dataset/wavs/ryz_ferenc/9788367950190_02_Hlasko_Palcie-ryz-kazdego-dnia_segment_55.wav|zˌatʃɨnˈamɨ? dˈɔbʒɛ.|0|Zaczynamy? Dobrze.
+Dataset/wavs/urwisko_ferenc/05_segment_56.wav|ʃtʃɛɡˈulɲʲɛ f tˌaki ˈupaw. pˌɔvʲɛdʑˈawaɕ ɕˌikɔrskʲˈɛmu, na jˈakix pɔtstˈavax pɔnˈɔvɲʲɛ fʃtʃˈɛntɔ ɕlˈɛtstfɔ?|0|Szczególnie w taki upał. Powiedziałaś Sikorskiemu, na jakich podstawach ponownie wszczęto śledztwo?
+Dataset/wavs/czytadlo_ferenc/9788367950787_011_Konwicki_Czytadlo_segment_26.wav|kˈɔrsak stˈanɔw ˈɔstrɔ.|0|Korsak stanął ostro.
+Dataset/wavs/czytadlo_ferenc/9788367950787_004_Konwicki_Czytadlo_segment_217.wav|ɲʲˈɛma prɔblˈɛmu, mˈɔɡɛ tɕˈɛ pˈɔdvʲɛɕtɕ. dʑɛŋkˈujɛ, tɔ ɲʲˌɛdalˈɛkɔ, mˈuʃɛ zˈɛbratɕ mˈɨɕli.|0|Nie ma problemu, mogę cię podwieźć. Dziękuję, to niedaleko, muszę zebrać myśli.
+Dataset/wavs/urwisko_ferenc/01_segment_78.wav|xˈɛrman ɲʲɛ zˌamjɛʒˈawa ɡˈɔ pɔɡˈaɲʲatɕ. tʃɛkˈawa tɕɛrplˈivʲɛ.|0|Herman nie zamierzała go poganiać. Czekała cierpliwie.
+Dataset/wavs/urwisko_ferenc/14_segment_8.wav|ˌalɛ xˈɛrman pˌɔɡrɔ̃ʒˈɨwa ɕɛ w rˌɔzmɨɕlˈaɲʲax i ɲʲɛ ˌɔtpɔvʲɛdʑˈawa.|0|Ale Herman pogrążyła się w rozmyślaniach i nie odpowiedziała.
+Dataset/wavs/urwisko_ferenc/14_segment_34.wav|tʃɨ s tɔ̃ lˈipskɔ̃, ˈɔprutʃ lˌɛvandˈɔfskʲɛj, mʲˈɛʃkaw ktˈɔɕ jˈɛʃtʃɛ?|0|Czy z tą Lipską, oprócz Lewandowskiej, mieszkał ktoś jeszcze?
+Dataset/wavs/ryz_ferenc/9788367950190_03_Hlasko_Palcie-ryz-kazdego-dnia_segment_122.wav|pˈɔtʃɔw znˈɔɕitɕ pˈatʃki ˌɔbvʲɔ̃zˈanɛ ʃnˈurkʲɛm i ukwˈadatɕ jˈɛ na drɛvɲʲˈanɛj pɔtstˈaftsɛ.|0|Począł znosić paczki obwiązane sznurkiem i układać je na drewnianej podstawce.
+Dataset/wavs/ryz_ferenc/9788367950190_15_Hlasko_Palcie-ryz-kazdego-dnia_segment_113.wav|pʃˌɛtɕɛʃ muɕˈawɛʑ bˈɨtɕ tˈak tʃɨ ˈɔvaɡ zdrˈɔvɔ pˌɔɡruxɔtˈanɨ.|0|Przecież musiałeś być tak czy owak zdrowo pogruchotany.
+Dataset/wavs/urwisko_ferenc/18_segment_88.wav|pɔdˈɔbnɔ vˌɨwɔvʲˈɔnɔ z vˈiswɨ ʃtʃˈɔntki jˈakʲɛjɕ kɔbjˈɛtɨ ɔt pˈaru.|0|Podobno wyłowiono z Wisły szczątki jakiejś kobiety od paru.
+Dataset/wavs/lew_ferenc/06_segment_240.wav|ˌalɛ jˈuʃ nˈavɛt tɕˈi, ktˌuʒɨ ftʃˈɛɕɲʲɛj ˌɔɡlɔndˈali bˌɔmbardɔvˈaɲʲa z dˈaxuf, ɲˈitʃɨm pɔkˈazɨ fˌajɛrvˈɛrkuf, zbjɛɡˈajɔ̃ tˈɛraz dɔ sxrˈɔnuf.|0|Ale już nawet ci, którzy wcześniej oglądali bombardowania z dachów, niczym pokazy fajerwerków, zbiegają teraz do schronów.
+Dataset/wavs/czytadlo_ferenc/9788367950787_010_Konwicki_Czytadlo_segment_397.wav|mɲʲˈɛjʃa s tˈɨm, jˈak mˈavʲaw bˌɨwɨ pˌɔtkɔmˈisaʃ kˈɔrsak.|0|Mniejsza z tym, jak mawiał były podkomisarz Korsak.
+Dataset/wavs/urwisko_ferenc/03_segment_130.wav|i pˈɔvʲɛm pˈaɲi, ʒˈɛ jɛst f tɛj muzˈɨtsɛ tsˈɔɕ, tsˈɔ mɲʲɛ pɔtɕˈɔŋɡa.|0|I powiem pani, że jest w tej muzyce coś, co mnie pociąga.
+Dataset/wavs/czytadlo_ferenc/9788367950787_002_Konwicki_Czytadlo_segment_121.wav|ʒˈɛbɨ ˌuratˈɔvatɕ ɕɛ i pʃˈɛʒɨtɕ, pɔtʃˈɛbna jɛzd dˌɨstsɨplˈina, ˌawtɔrˈɨtɛt.|0|Żeby uratować się i przeżyć, potrzebna jest dyscyplina, autorytet.
+Dataset/wavs/urwisko_ferenc/09_segment_113.wav|jˈarɔɲ ɔdvrˈutɕiw ɕɛ dɔ ɲʲˈɛj i pɔwˈɔʒɨw dwˈɔɲ na klˈamtsɛ.|0|Jaroń odwrócił się do niej i położył dłoń na klamce.
+Dataset/wavs/ryz_ferenc/9788367950190_10_Hlasko_Palcie-ryz-kazdego-dnia_segment_30.wav|pʃˌɛtɕɛʃ swˈɨʃɨ pˈaɲi, ʒˈɛ ja mˈuvjɛ z aktsˈɛntɛm.|0|Przecież słyszy pani, że ja mówię z akcentem.
+Dataset/wavs/urwisko_ferenc/36_segment_160.wav|— muvˈiwam jˈuʃ. xtɕˈawa, ʒˈɛbɨ pˈaɲi mʲˈawa ɡdʑˈɛ mʲˈɛʃkatɕ.|0|— Mówiłam już. Chciała, żeby pani miała gdzie mieszkać.
+Dataset/wavs/urwisko_ferenc/62_segment_25.wav|a xaŋɡʲˈɛrski zdˈawa na dʑˌɛvʲɛɲdʑˈɛɕɔnt kˈilka.|0|a hangierski zdała na dziewięćdziesiąt kilka.
+Dataset/wavs/czytadlo_ferenc/9788367950787_009_Konwicki_Czytadlo_segment_148.wav|pˌɔwɔʒˈɨwɛm ɕɛ na sfˈɔim tɕˈɔŋɡlɛ ɲʲˌɛzaswˈanɨm wˈuʃku.|0|Położyłem się na swoim ciągle niezasłanym łóżku.
+Dataset/wavs/urwisko_ferenc/46_segment_114.wav|ˌalɛ tˌɨlkɔ s kˌɔŋkrɛtˈami, ftˈɛdɨ pɔmˈɨɕlɛ.|0|Ale tylko z konkretami, wtedy pomyślę.
+Dataset/wavs/czytadlo_ferenc/9788367950787_004_Konwicki_Czytadlo_segment_270.wav|pˈɛwnɨm jˈakixɕ ɲʲɛjˈasnɨx pɔɡrˈuʒɛk.|0|Pełnym jakichś niejasnych pogróżek.
+Dataset/wavs/czytadlo_ferenc/9788367950787_004_Konwicki_Czytadlo_segment_4.wav|ˌɔtsɨkˈawɨ mɲʲɛ vɨbˈuxɨ ɕmʲˈɛxu na kˌɔrɨtˈaʒu.|0|Ocykały mnie wybuchy śmiechu na korytarzu.
+Dataset/wavs/czytadlo_ferenc/9788367950787_006_Konwicki_Czytadlo_segment_376.wav|vˌɨmɨɕlˈiliɕmˌɨ pˈɛvʲɛn spˈɔsup na zˌaɡwuʃˈaɲʲɛ vʲɛtʃnˈɛɡɔ ɲʲˌɛpɔkˈɔju.|0|Wymyśliliśmy pewien sposób na zagłuszanie wiecznego niepokoju.
+Dataset/wavs/ryz_ferenc/9788367950190_10_Hlasko_Palcie-ryz-kazdego-dnia_segment_73.wav|tɔ jɛst mɔdlˈitfa, tɨ skˌurvɨsˈɨnu, pɔvʲˈɛdʑaw ��andɛrsɔn.|0|To jest modlitwa, ty skurwysynu, powiedział Anderson.
+Dataset/wavs/czytadlo_ferenc/9788367950787_010_Konwicki_Czytadlo_segment_30.wav|i prˌɛzɛntˈɔvaw ɕɛ tˈɛras prˈavʲɛ ˌurɔtʃˈɨɕtɕɛ.|0|I prezentował się teraz prawie uroczyście.
+Dataset/wavs/urwisko_ferenc/43_segment_27.wav|pˈana zdˈaɲʲɛm mˈɔɡwa zapˈɔmɲʲɛtɕ ɔ dʑˈɛtsku i zɔstˈavitɕ jˈɛ f sˌamɔxˈɔdʑɛ?|0|Pana zdaniem mogła zapomnieć o dziecku i zostawić je w samochodzie?
+Dataset/wavs/urwisko_ferenc/18_segment_81.wav|kʲˈɛdɨ xudʑˈiɲski fʃˈɛd dɔ kˌɔrɨtˈaʒa, ɔtstˈaviw tˈɛtʃkɛ i pˌɔtsawˈɔvaw kɔbjˈɛtɛ w ˈusta, a nastˈɛmpɲʲɛ pˈɔdɲʲus xwˈɔptsa.|0|Kiedy Chudziński wszedł do korytarza, odstawił teczkę i pocałował kobietę w usta, a następnie podniósł chłopca.
+Dataset/wavs/ryz_ferenc/9788367950190_06_Hlasko_Palcie-ryz-kazdego-dnia_segment_275.wav|mɛ̃ʃtʃˈɨzna ɔdvrˈutɕiw ɕɛ ɡvawtˈɔvɲʲɛ, pˈatʃɨw pʃˈɛs xfˈilɛ na ɲʲˈɛɡɔ, a pˈɔtɛm ˈɔtʃɛd.|0|Mężczyzna odwrócił się gwałtownie, patrzył przez chwilę na niego, a potem odszedł.
+Dataset/wavs/urwisko_ferenc/08_segment_27.wav|stˈawɨ bɨvˈalɛts ˈimprɛs, kturˌɛɡɔ pˈartja prˌɔpaɡɔvˈawa xʃˌɛɕtɕijˈaɲskʲɛ vartˈɔɕtɕi.|0|Stały bywalec imprez, którego partia propagowała chrześcijańskie wartości.
+Dataset/wavs/urwisko_ferenc/19_segment_22.wav|ˌalɛ skˈɔrɔ vrakˈatɕɛ dɔ ɕlˈɛtstfa, muɕˈitɕɛ mʲˈɛtɕ kɔŋkrˈɛtnɨ pˈɔvut.|0|Ale skoro wracacie do śledztwa, musicie mieć konkretny powód.
+Dataset/wavs/lew_ferenc/08_segment_412.wav|zasˈɔbnɔʑdʑ zwˈuʃ pɔd bɨtˈɔmjɛm mˈa najvˈɨʃʃɛ pˌaramˈɛtrɨ f tsˈawɛj ɲʲˈɛtstsɛ vɛŋɡlˈɔvɛj.|0|Zasobność złóż pod Bytomiem ma najwyższe parametry w całej niecce węglowej.
+Dataset/wavs/ryz_ferenc/9788367950190_09_Hlasko_Palcie-ryz-kazdego-dnia_segment_120.wav|ɕˈɛdʑaw vʲˈɛndʑ w ɡˈurax rˈazɛm zˈɛ sfˈɔim sˈɨnɛm, ktˌurɨ mʲˈaw ɲʲɛvʲˈɛlɛ pˈɔnad dʑˈɛɕɛɲtɕ lˈat.|0|Siedział więc w górach razem ze swoim synem, który miał niewiele ponad dziesięć lat.
+Dataset/wavs/czytadlo_ferenc/9788367950787_010_Konwicki_Czytadlo_segment_108.wav|ˌɔbɛjʒˈawɛm fɔrsˈɨtsjɛ.|0|Obejrzałem forsycję.
+Dataset/wavs/ryz_ferenc/9788367950190_05_Hlasko_Palcie-ryz-kazdego-dnia_segment_238.wav|a ˈɛstɛr ˌɔdvinˈɛwa prˈavɨ rˈɛŋkaf jˈɛɡɔ zrudʑˈawɛj kˌanadˈɨjki.|0|A Ester odwinęła prawy rękaw jego zrudziałej Kanadyjki.
+Dataset/wavs/ryz_ferenc/9788367950190_12_Hlasko_Palcie-ryz-kazdego-dnia_segment_21.wav|tɛn tʃwˈɔvʲɛk ɲʲɛ ˈumjaw ʒˈɨdʑ bˈɛs latˈaɲʲa i bˈɨw jˈuʃ mˈartfɨ. nadwˈuɡɔ pʃˈɛttɛm ɲˈim ɕɛ zˈabʲiw.|0|Ten człowiek nie umiał żyć bez latania i był już martwy. Nadługo przedtem nim się zabił.
+Dataset/wavs/czytadlo_ferenc/9788367950787_006_Konwicki_Czytadlo_segment_312.wav|vʲˈɛm, ʒˈɛ jˈɛdna z nˈaʃɨx wɔntʃɲˈitʃɛk mʲˈɛʃka w ɡdˈaɲsku.|0|Wiem, że jedna z naszych łączniczek mieszka w Gdańsku.
+Dataset/wavs/ryz_ferenc/9788367950190_15_Hlasko_Palcie-ryz-kazdego-dnia_segment_4.wav|lˈɛdʒ dˈɛʒdʒ vrˈutɕiw znˈɔvu.|0|Lecz deszcz wrócił znowu.
+Dataset/wavs/ryz_ferenc/9788367950190_03_Hlasko_Palcie-ryz-kazdego-dnia_segment_306.wav|tʃˌɨtajˈɔntsɨ pʃˈɛrvaw i pɔvʲˈɛdʑaw pɔ xfˈili, f pʃɨʃwˈɔɕtɕi?|0|Czytający przerwał i powiedział po chwili, w przyszłości?
+Dataset/wavs/czytadlo_ferenc/9788367950787_003_Konwicki_Czytadlo_segment_386.wav|pˈan pɔzvˈɔli, ʒˈɛɡw dɔ mɲʲɛ i pɔkˈazaw ˌuvalˈanɔ̃ w ˌatramˈɛɲtɕɛ rˈɛŋkɔ̃ nˈaʃɛ zˌardʑɛvʲˈawɛ dʒvˈi.|0|Pan pozwoli, rzekł do mnie i pokazał uwalaną w atramencie ręką nasze zardzewiałe drzwi.
+Dataset/wavs/czytadlo_ferenc/9788367950787_001_Konwicki_Czytadlo_segment_111.wav|tsɔdʑˈɛnnɨm smˈɔɡʲɛm, rɔzmˈɨtɕɛm ˌɔddalˈɛɲʲa, ʒˈɛ pˌɔmɨɕlˈawɛm...|0|Codziennym smogiem, rozmyciem oddalenia, że pomyślałem...
+Dataset/wavs/ryz_ferenc/9788367950190_03_Hlasko_Palcie-ryz-kazdego-dnia_segment_419.wav|tˈak tɔ vɨɡlˈɔnda.|0|Tak to wygląda.

Data/normalize_log.csv

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:705c026e32b38f0151b35b10b91640f000c529c28ad8f4cc0d8fde960087788e
+size 10694681

Data/syspin.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63de2cd14152736cf300d73810def2c439f10caed73389448c64479db5920c83
+size 11781738

Data/texts/check.py

@@ -0,0 +1,102 @@
+#!/usr/bin/env python3
+import glob
+import os
+
+# ---------------- CONFIG ----------------
+INPUT_GLOB = "*.txt"       # which files to check (e.g. "*.txt", "data_*.txt")
+EXPECT_WAV_EXT = True      # set False if path doesn't have to end with .wav
+# ----------------------------------------
+
+
+def check_line_format(line: str, line_num: int, filename: str):
+    """
+    Check if a line matches: path|phonemes|speaker_id|language_id|rawtext
+    Returns None if OK, or an error message string if NOT OK.
+    """
+    stripped = line.rstrip("\n")
+
+    # Ignore empty lines
+    if not stripped.strip():
+        return None
+
+    # Ignore comment lines (optional)
+    if stripped.lstrip().startswith("#"):
+        return None
+
+    parts = stripped.split("|")
+
+    if len(parts) < 5:
+        return f"{filename}:{line_num}: expected >=5 fields, found {len(parts)} → {stripped!r}"
+
+    path = parts[0].strip()
+    phonemes = parts[1].strip()
+    speaker_id_str = parts[2].strip()
+    lang_id_str = parts[3].strip()
+    rawtext = "|".join(parts[4:]).strip()  # in case rawtext itself had pipes in future
+
+    if not path:
+        return f"{filename}:{line_num}: empty path field"
+
+    if EXPECT_WAV_EXT and not path.lower().endswith(".wav"):
+        return f"{filename}:{line_num}: path does not end with .wav → {path!r}"
+
+    # speaker_id must be int
+    try:
+        int(speaker_id_str)
+    except ValueError:
+        return f"{filename}:{line_num}: speaker_id not an int → {speaker_id_str!r}"
+
+    # language_id must be int
+    try:
+        int(lang_id_str)
+    except ValueError:
+        return f"{filename}:{line_num}: language_id not an int → {lang_id_str!r}"
+
+    if not rawtext:
+        return f"{filename}:{line_num}: empty rawtext field"
+
+    # phonemes we just require to exist (can be empty if you allow that)
+    # if you *require* non-empty phonemes, uncomment:
+    # if not phonemes:
+    #     return f"{filename}:{line_num}: empty phonemes field"
+
+    return None  # all good
+
+
+def main():
+    files = sorted(glob.glob(INPUT_GLOB))
+    if not files:
+        print("⚠️  No files matched pattern:", INPUT_GLOB)
+        return
+
+    print(f"📂 Checking {len(files)} file(s): {files}")
+
+    total_errors = 0
+    for fname in files:
+        errors_for_file = []
+        with open(fname, "r", encoding="utf-8") as f:
+            for i, line in enumerate(f, start=1):
+                err = check_line_format(line, i, fname)
+                if err:
+                    errors_for_file.append(err)
+
+        if errors_for_file:
+            total_errors += len(errors_for_file)
+            print(f"\n❌ {fname}: {len(errors_for_file)} line(s) with format issues")
+            # Show first few errors to avoid spam
+            for msg in errors_for_file[:10]:
+                print("   ", msg)
+            if len(errors_for_file) > 10:
+                print(f"   ... and {len(errors_for_file) - 10} more")
+        else:
+            print(f"✅ {fname}: all lines OK")
+
+    if total_errors == 0:
+        print("\n🎉 All checked files are in correct format.")
+    else:
+        print(f"\n⚠️  Done. Total problematic lines across all files: {total_errors}")
+
+
+if __name__ == "__main__":
+    main()
+

Data/texts/cor.py

@@ -0,0 +1,139 @@
+#!/usr/bin/env python3
+import os
+
+INPUT_FILE = "syspin.txt"
+OUTPUT_FILE = "syspin_fixed.txt"
+DEFAULT_LANG_ID = 2           # Hindi language id
+PATH_PREFIX = "syspin/"       # what a “real” record line starts with
+
+
+def join_logical_lines(path_prefix=PATH_PREFIX):
+    """
+    Join wrapped lines so each logical record becomes a single string.
+
+    Any line starting with `path_prefix` (e.g. 'syspin/') starts a new record.
+    Any following non-empty line that does NOT start with that prefix is
+    treated as continuation of the previous record's raw text.
+    """
+    logical = []
+    current = None
+
+    with open(INPUT_FILE, "r", encoding="utf-8") as f:
+        for raw in f:
+            stripped = raw.rstrip("\n")
+            if not stripped.strip():
+                continue  # skip empty lines
+
+            if stripped.startswith(path_prefix):
+                # Start of a new record
+                if current is not None:
+                    logical.append(current)
+                current = stripped
+            else:
+                # Continuation of previous record's text
+                if current is None:
+                    # stray line, ignore or keep separately if you want
+                    continue
+                current += " " + stripped.strip()
+
+        if current is not None:
+            logical.append(current)
+
+    return logical
+
+
+def fix_record(line: str) -> str:
+    """
+    Parse a single logical record and ensure it has:
+        path|phonemes|speaker|lang|rawtext...
+
+    Handles:
+    - old 4-field format: path|phonemes|speaker|rawtext...
+    - new 5-field format: path|phonemes|speaker|lang|rawtext...
+
+    Rawtext can contain '|' and spaces.
+    """
+    line = line.strip()
+    if not line:
+        return line
+
+    # 1) Split off path
+    try:
+        path, rest = line.split("|", 1)
+    except ValueError:
+        # No '|' at all, just return as-is
+        return line
+
+    path = path.strip()
+
+    # 2) Split off phonemes
+    try:
+        phonemes, rest2 = rest.split("|", 1)
+    except ValueError:
+        return line  # malformed, leave as-is
+
+    # 3) Try to parse speaker, lang, rawtext with controlled splits
+    # rest2 is "speaker|lang|rawtext..." or "speaker|rawtext..."
+    parts = rest2.split("|", 2)
+
+    # Case A: we have at least speaker, lang, rawtext
+    if len(parts) == 3:
+        speaker_str = parts[0].strip()
+        lang_or_raw = parts[1].strip()
+        raw_rest = parts[2]  # may contain more '|'
+
+        # speaker must be int
+        try:
+            int(speaker_str)
+        except ValueError:
+            return line  # not our expected format
+
+        # If lang_or_raw is int, this is already 5-field format
+        try:
+            int(lang_or_raw)
+            lang_str = lang_or_raw
+            rawtext = raw_rest
+            return f"{path}|{phonemes}|{speaker_str}|{lang_str}|{rawtext}"
+        except ValueError:
+            # This is actually 4-field, but rawtext itself had '|'
+            # Original shape: speaker|raw_part1|raw_rest
+            # So we reassemble rawtext:
+            rawtext = lang_or_raw + "|" + raw_rest
+            lang_str = str(DEFAULT_LANG_ID)
+            return f"{path}|{phonemes}|{speaker_str}|{lang_str}|{rawtext}"
+
+    # Case B: we only have "speaker|rawtext..."
+    elif len(parts) == 2:
+        speaker_str = parts[0].strip()
+        rawtext = parts[1]
+
+        try:
+            int(speaker_str)
+        except ValueError:
+            return line
+
+        lang_str = str(DEFAULT_LANG_ID)
+        return f"{path}|{phonemes}|{speaker_str}|{lang_str}|{rawtext}"
+
+    # Anything else: leave unchanged
+    return line
+
+
+def main():
+    print(f"🔍 Joining logical lines from {INPUT_FILE}...")
+    logical_lines = join_logical_lines()
+
+    print(f"📦 Found {len(logical_lines)} logical records. Fixing format...")
+    fixed = [fix_record(l) for l in logical_lines]
+
+    with open(OUTPUT_FILE, "w", encoding="utf-8") as f:
+        for l in fixed:
+            f.write(l + "\n")
+
+    print(f"✅ Wrote fixed file to {OUTPUT_FILE}")
+    print("👉 Now run your check script on syspin_fixed.txt instead of syspin.txt")
+
+
+if __name__ == "__main__":
+    main()
+

Data/texts/fin.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b6553b3035a803ba52283e8c1b1e02df58c08906b1f530b37a7f0741861c470
+size 36893682

Data/texts/merge.py

@@ -0,0 +1,46 @@
+#!/usr/bin/env python3
+import os
+import glob
+
+# ---------------- CONFIG ----------------
+INPUT_GLOB = "*.txt"                # all text files in current directory
+OUTPUT_FILE = "merged_sorted.txt"   # output filename
+LANGUAGE_COL_INDEX = 3              # 0-based index (4th column = language id)
+# ----------------------------------------
+
+def parse_line(line):
+    """Split and return (lang_id, line) if valid."""
+    parts = line.strip().split("|")
+    if len(parts) > LANGUAGE_COL_INDEX:
+        try:
+            lang_id = int(parts[LANGUAGE_COL_INDEX])
+            return lang_id, line.strip()
+        except ValueError:
+            pass
+    return None, line.strip()
+
+def main():
+    # Gather all .txt files
+    files = sorted(glob.glob(INPUT_GLOB))
+    print(f"📂 Found {len(files)} text files: {files}")
+
+    all_lines = []
+    for file in files:
+        with open(file, "r", encoding="utf-8") as f:
+            for line in f:
+                if line.strip():
+                    lang_id, full_line = parse_line(line)
+                    all_lines.append((lang_id, full_line))
+
+    # Sort by language ID (then by filename order for consistency)
+    all_lines.sort(key=lambda x: (x[0], x[1]))
+
+    with open(OUTPUT_FILE, "w", encoding="utf-8") as f:
+        for _, line in all_lines:
+            f.write(line + "\n")
+
+    print(f"✅ Merged and sorted {len(all_lines)} lines into {OUTPUT_FILE}")
+
+if __name__ == "__main__":
+    main()
+

Data/train_3.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a91f40aa64b12431823c2046ff835e8a5bb7e624c47b5915547b6a2558cf6bc
+size 47346672

Data/train_3_filtered.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd29f1d22ef76f02d7843fd6ffbe46ee12c38e52e4f6f0614a6f989c8029452d
+size 35290704

Data/train_3_filtered_1.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e204501375a046beda88812643e222be5c49346ea690a5215db73a9838d3c188
+size 35253822

Data/train_3_filtered_dur.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2930e002060edbac2eea683909074d4270a02e66e759ccd7830e73c256efecca
+size 32913633

Data/train_3_filtered_dur_fixed.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1f1c6cd7d04573ea56fbbd24792e28c9716ace5bc4b7ec2b991d8a4ac7ac57f
+size 33069989

Data/train_3_filtered_dur_fixed_copy.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fad71ba31150e12bb27836601326157ac6791b227ef7104a0d26d6a5cb3e9a2c
+size 30886753

Data/train_3_filtered_dur_fixed_speakers.txt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f3fb1ccb7717292f5f226a0747f211e24bcb435bfbe9798066aa853dfcb203e
+size 33069989