| # Complete EMA model; inference only. | |
| model: | |
| name: WaveGen_Flash | |
| tokenizer: pinyin | |
| backbone: DiT | |
| cfm: | |
| prediction: x_pred | |
| t_eps: 0.02 | |
| latents_scale: 9.0 | |
| arch: | |
| dim: 1152 | |
| depth: 24 | |
| heads: 16 | |
| dropout: 0.0 | |
| ff_mult: 4 | |
| text_dim: 512 | |
| text_mask_padding: true | |
| qk_norm: null | |
| conv_layers: 4 | |
| pe_attn_head: null | |
| attn_backend: torch | |
| attn_mask_enabled: false | |
| checkpoint_activations: false | |
| mel_channels: 80 | |
| use_mel_information_bottleneck: true | |
| mel_bottleneck_dim: 32 | |
| mel_bottleneck_dropout: 0.2 # Disabled explicitly by inference_bottleneck_dropout below. | |
| mel_bottleneck_zero_init: false | |
| use_audio_proj: true | |
| audio_proj_dim: 1024 | |
| audio_proj_hidden: 2048 | |
| waveform: | |
| target_sample_rate: 24000 | |
| wav_frame_len: 480 | |
| inference: | |
| num_steps: 4 | |
| sampling_method: euler | |
| time_shift: 3.0 | |
| # Already shifted: t / (t + 3 * (1 - t)), from [0, .25, .5, .75, 1]. | |
| # Use these exact evaluation times; do not apply another time transform. | |
| time_grid: [0.0, 0.1, 0.25, 0.5, 1.0] | |
| cfg_strength: 0.0 | |
| inference_bottleneck_dropout: 0.0 | |
| waveform_temperature: 1.0 | |
| # Initial Mel noise only; use 1.0 to reproduce the previous setting. | |
| mel_temperature: 0.0 | |
| compute_final_mel: false | |
| max_duration_seconds: 30.0 | |