ICTNLP/indextts2_decoder

IndexTTS2-derived waveform decoder for FreezeEmpath

This package converts FreezeEmpath speech tokens into 22,050 Hz mono waveforms, using components from IndexTTS2, Amphion/MaskGCT, and NVIDIA BigVGAN.

FreezeEmpath model · Code and instructions · Paper

Contents and use

• config.yaml: semantic codec and flow-matching configuration.
• semantic_codec.safetensors: semantic codec weights.
• s2mel.pth: flow-matching model weights.
• bigvgan_generator.pt: vocoder weights.
• style.pt, prompt_condition.pt, ref_mel.pt: fixed voice conditions used by the released example.

Use the FreezeEmpath GitHub code and its bundled BigVGAN architecture configuration:

python inference/codes2wav.py \
  --cfg_path checkpoints/indextts2_codes2wav/config.yaml \
  --model_dir checkpoints/indextts2_codes2wav \
  --input_file outputs/answers.jsonl \
  --output_dir outputs/wav --seed 42

The input must contain prediction_units speech tokens generated by inference/infer.py --s2s; this package is not a standalone text-to-speech interface. The example uses a fixed voice, not user-selected speaker prompts.

License

This IndexTTS2-derived package is subject to the bilibili Model Use License Agreement. Individual BigVGAN and Amphion components retain their upstream MIT notices. See third-party notices.

Any modifications made to the original model in this Derivative Work are not endorsed, warranted, or guaranteed by the original right-holder of the original model, and the original right-holder disclaims all liability related to this Derivative Work.