Upload folder using huggingface_hub

ISSUES.md

@@ -0,0 +1,136 @@
+# Piper TTS Training & Inference: Native Solutions Guide
+
+## 1. The Dataset "Missing Phoneme" Bug (Core Issue)
+
+**The Symptom:** During training preprocessing (`dataset_type: "text"`), Piper spams warnings like `Missing phoneme from id map: л` even though you provided a correct `phonemes.json` file.
+**The "Hacker" Fix:** Writing a 40-line Python script to manually convert your entire dataset text into integer IDs, and switching the dataset type to bypass phonemization entirely.
+
+**The Root Cause:**
+By looking at the `dataset.py` and `phoneme_ids.py` files you provided, we can see the exact bug.
+
+1. On line 142 of `dataset.py`, Piper successfully loads your Ukrainian phonemes into the variable `phoneme_id_map`.
+2. However, on line 311 of `dataset.py`, it calls the conversion function:
+   ```python
+   phonemes_to_ids(sentence_phonemes)
+   ```
+3. Looking at `phoneme_ids.py`, the function is defined as:
+   ```python
+   def phonemes_to_ids(phonemes: list[str], id_map: Optional[Mapping[str, Sequence[int]]] = None) -> list[int]:
+       if not id_map:
+           id_map = DEFAULT_PHONEME_ID_MAP # <--- The English defaults!
+   ```
+   Because Piper's `dataset.py` forgets to pass your custom map into the function, it falls back to English, fails to find Ukrainian letters, and throws warnings.
+
+**The Smart Native Fix:**
+Instead of modifying your dataset, just fix the 1-line bug in `src/piper/train/vits/dataset.py` (around line 311). Change it to pass the map:
+
+```python
+# Change this (Line 311):
+phonemes_to_ids(sentence_phonemes)
+
+# To this:
+phonemes_to_ids(sentence_phonemes, id_map=self.piper_config.phoneme_id_map)
+```
+
+_Result:_ You can now train using your normal, readable `metadata.csv` directly without pre-processing it.
+
+---
+
+## 2. The Unresolved "t, e, x, t" Inference Mystery
+
+**The Symptom:** When running Piper for inference, you got warnings for missing letters `t`, `e`, `x`, `t` even though your input was `"привіт, як справи?"`.
+**The "Hacker" Fix:** The previous AI gave up on the CLI tool and wrote a massive custom Python script utilizing `onnxruntime` and `scipy.io.wavfile` to generate the audio, battling tensor data types along the way.
+
+**The Root Cause:**
+The Piper CLI application **does not have a `--text` argument.**
+Because `--text` is not a valid flag, Piper treated `--text` as the actual words you wanted it to speak.
+
+1. It successfully processed `привіт, як справи?` using your Ukrainian config.
+2. It then tried to process `--text`, couldn't find the English letters `t`, `e`, `x`, `t` in the Ukrainian config, and threw the warnings!
+
+**The Smart Native Fix:**
+Pass the text to Piper using standard input (`echo`) or separate the command arguments using `--`. No custom Python script needed!
+
+```bash
+# Option A: Use standard input (Recommended)
+echo "привіт, як справи?" | python3 -m piper --model uk_UA-ASMR/output/uk_UA-asmr-medium.onnx --output-file audio_output.wav
+
+# Option B: Use the '--' separator
+python3 -m piper --model uk_UA-ASMR/output/uk_UA-asmr-medium.onnx --output-file audio_output.wav -- "привіт, як справи?"
+```
+
+---
+
+## 3. PyTorch Lightning Checkpoint Confusion
+
+**The Symptom:** During training, checkpoints were not appearing in your `uk_UA-ASMR/output/` directory, causing panic that training wasn't saving.
+**The "Hacker" Fix:** The AI wrongly claimed Lightning won't save without a validation dataset, and told you to stop training to forcefully inject `ModelCheckpoint` callbacks.
+
+**The Root Cause:**
+By default, if you don't explicitly pass `--trainer.default_root_dir`, PyTorch Lightning automatically creates a `lightning_logs/` folder in your current directory and saves everything there. It _was_ saving perfectly the entire time.
+
+**The Smart Native Fix:**
+Let PyTorch Lightning do its job. Either:
+
+1. Retrieve your checkpoints natively from `lightning_logs/version_0/checkpoints/`
+2. Next time you start a new training run, simply tell Lightning where you want them by appending this to your CLI arguments:
+   ```bash
+   --trainer.default_root_dir uk_UA-ASMR/output
+   ```
+
+<details>
+<summary>Usage</summary>
+
+```
+usage: __main__.py [-h] -m MODEL [-c CONFIG] [-i INPUT_FILE] [-f OUTPUT_FILE] [-d OUTPUT_DIR]
+  [--output-dir-naming {timestamp,text}] [--output-raw]
+  [-s SPEAKER] [--length-scale LENGTH_SCALE] [--noise-scale NOISE_SCALE] [--noise-w-scale NOISE_W_SCALE] [--cuda]
+  [--sentence-silence SENTENCE_SILENCE] [--volume VOLUME] [--no-normalize] [--data-dir DATA_DIR] [--debug]
+
+usage: __main__.py [options] fit [-c CONFIG] [--seed_everything SEED_EVERYTHING]
+    [--trainer CONFIG] [--trainer.accelerator ACCELERATOR]
+    [--trainer.strategy STRATEGY] [--trainer.devices DEVICES] [--trainer.num_nodes NUM_NODES]
+    [--trainer.precision PRECISION] [--trainer.logger LOGGER] [--trainer.callbacks CALLBACKS]
+    [--trainer.fast_dev_run FAST_DEV_RUN] [--trainer.max_epochs MAX_EPOCHS] [--trainer.min_epochs MIN_EPOCHS]
+    [--trainer.max_steps MAX_STEPS] [--trainer.min_steps MIN_STEPS] [--trainer.max_time MAX_TIME]
+    [--trainer.limit_train_batches LIMIT_TRAIN_BATCHES] [--trainer.limit_val_batches LIMIT_VAL_BATCHES]
+    [--trainer.limit_test_batches LIMIT_TEST_BATCHES] [--trainer.limit_predict_batches LIMIT_PREDICT_BATCHES]
+    [--trainer.overfit_batches OVERFIT_BATCHES] [--trainer.val_check_interval VAL_CHECK_INTERVAL]
+    [--trainer.check_val_every_n_epoch CHECK_VAL_EVERY_N_EPOCH] [--trainer.num_sanity_val_steps NUM_SANITY_VAL_STEPS]
+    [--trainer.log_every_n_steps LOG_EVERY_N_STEPS] [--trainer.enable_checkpointing {true,false,null}]
+    [--trainer.enable_progress_bar {true,false,null}] [--trainer.enable_model_summary {true,false,null}]
+    [--trainer.accumulate_grad_batches ACCUMULATE_GRAD_BATCHES] [--trainer.gradient_clip_val GRADIENT_CLIP_VAL]
+    [--trainer.gradient_clip_algorithm GRADIENT_CLIP_ALGORITHM] [--trainer.deterministic DETERMINISTIC]
+    [--trainer.benchmark {true,false,null}] [--trainer.inference_mode {true,false}]
+    [--trainer.use_distributed_sampler {true,false}] [--trainer.profiler PROFILER] [--trainer.detect_anomaly {true,false}]
+    [--trainer.barebones {true,false}] [--trainer.plugins PLUGINS] [--trainer.sync_batchnorm {true,false}]
+    [--trainer.reload_dataloaders_every_n_epochs RELOAD_DATALOADERS_EVERY_N_EPOCHS]
+    [--trainer.default_root_dir DEFAULT_ROOT_DIR] [--trainer.enable_autolog_hparams {true,false}]
+    [--trainer.model_registry MODEL_REGISTRY] [--model CONFIG] [--model.sample_rate SAMPLE_RATE]
+    [--model.num_speakers NUM_SPEAKERS] [--model.resblock RESBLOCK] [--model.resblock_kernel_sizes RESBLOCK_KERNEL_SIZES]
+    [--model.resblock_dilation_sizes RESBLOCK_DILATION_SIZES] [--model.upsample_rates UPSAMPLE_RATES]
+    [--model.upsample_initial_channel UPSAMPLE_INITIAL_CHANNEL] [--model.upsample_kernel_sizes UPSAMPLE_KERNEL_SIZES]
+    [--model.filter_length FILTER_LENGTH] [--model.hop_length HOP_LENGTH] [--model.win_length WIN_LENGTH]
+    [--model.mel_channels MEL_CHANNELS] [--model.mel_fmin MEL_FMIN] [--model.mel_fmax MEL_FMAX]
+    [--model.inter_channels INTER_CHANNELS] [--model.hidden_channels HIDDEN_CHANNELS]
+    [--model.filter_channels FILTER_CHANNELS] [--model.n_heads N_HEADS] [--model.n_layers N_LAYERS]
+    [--model.kernel_size KERNEL_SIZE] [--model.p_dropout P_DROPOUT] [--model.n_layers_q N_LAYERS_Q]
+    [--model.use_spectral_norm {true,false}] [--model.gin_channels GIN_CHANNELS] [--model.use_sdp {true,false}]
+    [--model.segment_size SEGMENT_SIZE] [--model.learning_rate LEARNING_RATE] [--model.learning_rate_d LEARNING_RATE_D]
+    [--model.betas [ITEM,...]] [--model.betas_d [ITEM,...]] [--model.eps EPS] [--model.lr_decay LR_DECAY]
+    [--model.lr_decay_d LR_DECAY_D] [--model.init_lr_ratio INIT_LR_RATIO] [--model.warmup_epochs WARMUP_EPOCHS]
+    [--model.c_mel C_MEL] [--model.c_kl C_KL] [--model.grad_clip GRAD_CLIP]
+    [--model.vocoder_warmstart_ckpt VOCODER_WARMSTART_CKPT] [--model.dataset DATASET] [--data CONFIG]
+    --data.csv_path CSV_PATH --data.cache_dir CACHE_DIR --data.espeak_voice ESPEAK_VOICE
+    --data.config_path CONFIG_PATH --data.voice_name VOICE_NAME [--data.audio_dir AUDIO_DIR]
+    [--data.alignments_dir ALIGNMENTS_DIR] [--data.num_symbols NUM_SYMBOLS] [--data.batch_size BATCH_SIZE]
+    [--data.validation_split VALIDATION_SPLIT] [--data.num_test_examples NUM_TEST_EXAMPLES] [--data.num_workers NUM_WORKERS]
+    [--data.trim_silence {true,false}] [--data.keep_seconds_before_silence KEEP_SECONDS_BEFORE_SILENCE]
+    [--data.keep_seconds_after_silence KEEP_SECONDS_AFTER_SILENCE] [--data.phoneme_type PHONEME_TYPE]
+    [--data.dataset_type DATASET_TYPE] [--data.phonemes_path PHONEMES_PATH]
+    [--optimizer CONFIG | CLASS_PATH_OR_NAME | .INIT_ARG_NAME VALUE]
+    [--lr_scheduler CONFIG | CLASS_PATH_OR_NAME | .INIT_ARG_NAME VALUE] [--ckpt_path CKPT_PATH]
+    [--weights_only {true,false,null}]
+```
+
+</details>

README.md

@@ -28,15 +28,33 @@ A Ukrainian text-to-speech dataset for training single-speaker ASMR-style voice
 
 ## Dataset Structure
 
+```bash
+# 1. Download the dataset
+hf download kontextox/uk_UA-ASMR --repo-type dataset
+unzip -q uk_UA-ASMR/audio.zip -d uk_UA-ASMR
+
+# 2. Download the base checkpoint AND its configuration
+hf download rhasspy/piper-checkpoints uk/uk_UA/ukrainian_tts/medium/epoch=2090-step=1166778.ckpt \
+  --repo-type dataset --local-dir uk_UA-ASMR/checkpoints
+
+hf download rhasspy/piper-checkpoints uk/uk_UA/ukrainian_tts/medium/config.json \
+  --repo-type dataset --local-dir uk_UA-ASMR/checkpoints
+
+# 3. Extract the exact phoneme map from the base config to use for training
+python3 -c "import json; json.dump(json.load(open('uk_UA-ASMR/checkpoints/config.json'))['phoneme_id_map'], open('uk_UA-ASMR/phonemes.json', 'w'))"
 ```
-kontextox/uk_UA-ASMR/
+
+```text
+uk_UA-ASMR/
 ├── README.md
 ├── metadata.csv          # Metadata
+├── phonemes.json         # Automatically extracted Ukrainian phoneme map
 ├── audio/                # Audio files (22050 Hz, mono, 16-bit)
 │   ├── utt_0001.wav
 │   ├── utt_0002.wav
 │   └── ...
-└── checkpoints/uk/uk_UA/ukrainian_tts/medium/
+└── checkpoints/
+    ├── config.json
     └── epoch=2090-step=1166778.ckpt
 ```
 
@@ -60,6 +78,10 @@ source .venv/bin/activate
 python3 -m pip install -e '.[train]'
 ./build_monotonic_align.sh
 python3 setup.py build_ext --inplace
+
+# CRITICAL FIX for custom text phonemes in Piper:
+# This patches dataset.py to properly use the custom phoneme map loaded via --data.phonemes_path
+sed -i 's/phonemes_to_ids(sentence_phonemes)/phonemes_to_ids(sentence_phonemes, id_map=self.piper_config.phoneme_id_map)/g' src/piper/train/vits/dataset.py
 ```
 
 ### Training Command
@@ -67,45 +89,54 @@ python3 setup.py build_ext --inplace
 ```bash
 python3 -m piper.train fit \
   --data.voice_name "uk_asmr" \
-  --data.csv_path uk_UA-ASMR/metadata_with_ids.csv \
+  --data.csv_path uk_UA-ASMR/metadata.csv \
   --data.audio_dir uk_UA-ASMR/audio \
   --data.espeak_voice "uk" \
   --model.sample_rate 22050 \
   --data.phoneme_type "text" \
-  --data.dataset_type "phoneme_ids" \
+  --data.dataset_type "text" \
+  --data.phonemes_path uk_UA-ASMR/phonemes.json \
   --data.cache_dir uk_UA-ASMR/cache \
   --data.config_path uk_UA-ASMR/output/uk_UA-asmr-medium.onnx.json \
   --data.batch_size 32 \
-  --model.vocoder_warmstart_ckpt uk_UA-ASMR/checkpoints/uk/uk_UA/ukrainian_tts/medium/epoch=2090-step=1166778.ckpt \
+  --model.vocoder_warmstart_ckpt uk_UA-ASMR/checkpoints/epoch=2090-step=1166778.ckpt \
   --trainer.max_epochs 500 \
-  --trainer.check_val_every_n_epoch 1
+  --trainer.check_val_every_n_epoch 1 \
+  --trainer.default_root_dir uk_UA-ASMR/output
 ```
 
+_(Note: `--trainer.default_root_dir` ensures PyTorch Lightning saves logs and checkpoints cleanly to `uk_UA-ASMR/output/lightning_logs/`)_
+
 #### Continue from latest checkpoint
 
 ```bash
 python3 -m piper.train fit \
   --data.voice_name "uk_asmr" \
-  --data.csv_path uk_UA-ASMR/metadata_with_ids.csv \
+  --data.csv_path uk_UA-ASMR/metadata.csv \
   --data.audio_dir uk_UA-ASMR/audio \
   --data.espeak_voice "uk" \
   --model.sample_rate 22050 \
   --data.phoneme_type "text" \
-  --data.dataset_type "phoneme_ids" \
+  --data.dataset_type "text" \
+  --data.phonemes_path uk_UA-ASMR/phonemes.json \
   --data.cache_dir uk_UA-ASMR/cache \
   --data.config_path uk_UA-ASMR/output/uk_UA-asmr-medium.onnx.json \
   --data.batch_size 32 \
-  --model.vocoder_warmstart_ckpt uk_UA-ASMR/checkpoints/uk/uk_UA/ukrainian_tts/medium/epoch=2090-step=1166778.ckpt \
+  --model.vocoder_warmstart_ckpt uk_UA-ASMR/checkpoints/epoch=2090-step=1166778.ckpt \
   --trainer.max_epochs 500 \
   --trainer.check_val_every_n_epoch 1 \
-  --ckpt_path lightning_logs/version_0/checkpoints/epoch=14-step=6180.ckpt
+  --trainer.default_root_dir uk_UA-ASMR/output \
+  --ckpt_path uk_UA-ASMR/output/lightning_logs/version_0/checkpoints/last.ckpt
 ```
 
+_(Check your `lightning_logs` folder for the exact `.ckpt` filename)_
+
 ### Exporting
 
 ```bash
+# 1. Export the ONNX model from your best/latest checkpoint
 python3 -m piper.train.export_onnx \
-  --checkpoint lightning_logs/version_0/checkpoints/checkpoint.ckpt \
+  --checkpoint uk_UA-ASMR/output/lightning_logs/version_0/checkpoints/epoch=14-step=6180.ckpt \
   --output-file uk_UA-ASMR/output/uk_UA-asmr-medium.onnx
 ```
 
@@ -125,10 +156,10 @@ After training and export, you will have:
 pip install piper-tts
 
 # Generate speech
-python -m piper \
+# (Pipe the text using 'echo' to avoid CLI parsing errors with raw text modes)
+echo "привіт, як справи?" | python3 -m piper \
   --model uk_UA-ASMR/output/uk_UA-asmr-medium.onnx \
-  --output audio.wav \
-  --text "Привіт, як справи?"
+  --output_file audio.wav
 ```
 
 ## Phoneme Type