inference.yaml

# ################################
# ProTeVa Inference Configuration
# Simplified YAML for deployment
# ################################

# Basic settings
seed: 200
device: cpu  # Change to cuda if GPU available
sample_rate: 16000

# Output neurons (4 classes: blank, high, low, mid)
# Based on labelencoder.txt: 0=blank, 1=H, 2=B, 3=M
# Space (4) is added via post-processing
output_neurons: 4
blank_index: 0

# Number of prototypes
n_prototypes: 10

# Feature dimension from HuBERT
emb_dim: 768

# Encoder settings
rnn_layers: 2
rnn_neurons: 512

# Decoder settings
dnn_blocks: 2
dnn_neurons: 512

# Pitch decoder settings
dec_dnn_blocks: [1]
dec_dnn_neurons: [128]

# Activation function
activation: !name:torch.nn.LeakyReLU

# ============ MODULES ============

# HuBERT feature extractor
wav2vec2: !new:speechbrain.lobes.models.huggingface_transformers.hubert.HuBERT
   source: "Orange/SSA-HuBERT-base-60k"
   output_norm: True
   freeze: False
   save_path: whubert_checkpoint

# F0 extractor (requires custom module)
f0Compute: !new:modules.F0Extractor
    device: !ref <device>
    sample_rate: !ref <sample_rate>

# BiGRU Encoder
enc: !new:speechbrain.nnet.RNN.GRU
    input_shape: [null, null, !ref <emb_dim>]
    hidden_size: !ref <rnn_neurons>
    num_layers: !ref <rnn_layers>
    bidirectional: True
    dropout: 0.15

# VanillaNN Decoder
dec: !new:speechbrain.lobes.models.VanillaNN.VanillaNN
    input_shape: [null, null, 1024]  # 512 * 2 (bidirectional)
    activation: !ref <activation>
    dnn_blocks: !ref <dnn_blocks>
    dnn_neurons: !ref <dnn_neurons>

# Pitch Decoder (requires custom module)
pitch_dec: !new:modules.PitchDecoderLayer
    input_shape: [null, null, !ref <dnn_neurons>]
    dnn_blocks: !ref <dec_dnn_blocks>
    dnn_neurons: !ref <dec_dnn_neurons>

# Prototype Layer (requires custom module)
proto: !new:modules.PrototypeLayer
    n_prototypes: !ref <n_prototypes>
    latent_dims: !ref <dnn_neurons>

# Output linear layer
output_lin: !new:speechbrain.nnet.linear.Linear
    input_size: !ref <n_prototypes>
    n_neurons: !ref <output_neurons>
    bias: True

# Log softmax
log_softmax: !new:speechbrain.nnet.activations.Softmax
    apply_log: True

# Label encoder
label_encoder: !new:speechbrain.dataio.encoder.CTCTextEncoder

# ============ MODULES DICT ============

modules:
    wav2vec2: !ref <wav2vec2>
    enc: !ref <enc>
    dec: !ref <dec>
    pitch_dec: !ref <pitch_dec>
    proto: !ref <proto>
    output_lin: !ref <output_lin>

# Model container for all modules
model: !new:torch.nn.ModuleList
    - [!ref <enc>, !ref <dec>, !ref <proto>, !ref <output_lin>, !ref <pitch_dec>]

# ============ PRETRAINER ============
# This loads the trained checkpoints

pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
    loadables:
        model: !ref <model>
        wav2vec2: !ref <wav2vec2>
        tokenizer: !ref <label_encoder>
    paths:
        model: !ref <save_folder>/model.ckpt
        wav2vec2: !ref <save_folder>/wav2vec2.ckpt
        tokenizer: !ref <save_folder>/tokenizer.ckpt

# Save folder - Path is loaded from config.py
# To change checkpoint folder, update CHECKPOINT_FOLDER in config.py
save_folder: ./CKPT+2025-10-20+08-19-07+00