| # 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 |