---
language:
  - en
pipeline_tag: text-to-speech
license: other
license_name: apache-2.0-and-mit
license_link: https://huggingface.co/DOOD02/EditVoice/blob/main/LICENSE
tags:
  - speech-editing
  - text-to-speech
  - audio
---

# EditVoice — Official Model Weights

This is the official model weights repository for [EditVoice: Variable-Length Non-Autoregressive Zero-Shot TTS and Speech Editing with Edit Flows](https://arxiv.org/abs/2609.29889). It provides the inference checkpoints and Edit Flow initial distribution for zero-shot TTS and speech editing in noisy environments. The [official code repository](https://github.com/DHY02/EditVoice) contains the inference implementation, and audio examples are available in the [EditVoice demo](https://dhy02.github.io/editvoice-demo/).

## Files and Provenance

| File | Purpose | Source and license |
|---|---|---|
| `checkpoints/editvoice_lm.pt` | Edit Flow semantic model for TTS and speech editing | EditVoice; Apache-2.0 |
| `checkpoints/kimi-flow.pt` | Token-to-mel Flow Matching model fine-tuned for speech editing in noisy environments; paired with `checkpoints/kimi-bigvgan/` | EditVoice fine-tuning from CosyVoice2 `flow.pt`; Apache-2.0 |
| `data/empirical.pkl` | 6,561-class initial distribution $p_0$ used by Edit Flows | EditVoice training-data statistics; Apache-2.0 |
| `checkpoints/flow.pt` | Token-to-mel Flow Matching model for zero-shot TTS; paired with `checkpoints/hift.pt` | [CosyVoice2-0.5B](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B); Apache-2.0 |
| `checkpoints/hift.pt` | Vocoder for zero-shot TTS; paired with `checkpoints/flow.pt` | [CosyVoice2-0.5B](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B); Apache-2.0 |
| `checkpoints/speech_tokenizer_v2.onnx` | Speech tokenizer | [CosyVoice2-0.5B](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B); Apache-2.0 |
| `checkpoints/campplus.onnx` | Speaker embedding extractor | [CosyVoice2-0.5B](https://huggingface.co/FunAudioLLM/CosyVoice2-0.5B); Apache-2.0 |
| `checkpoints/kimi-bigvgan/config.json`, `checkpoints/kimi-bigvgan/model.pt` | Vocoder for speech editing in noisy environments; paired with `checkpoints/kimi-flow.pt` | `vocoder/` from [Kimi-Audio-7B-Instruct](https://huggingface.co/moonshotai/Kimi-Audio-7B-Instruct); MIT |

The `ttsfrd` wheel and its English resources are not included; obtain them separately from an authorized source.

## Download and Use

With the Hugging Face CLI installed, download the files and verify their checksums:

```bash
hf download DOOD02/EditVoice --local-dir model_assets
(cd model_assets && sha256sum -c SHA256SUMS)
```

For `editvoice-tts`, pass `model_assets/checkpoints/editvoice_lm.pt`, `flow.pt`, `hift.pt`, `speech_tokenizer_v2.onnx`, `campplus.onnx`, and `model_assets/data/empirical.pkl` to their corresponding checkpoint and resource arguments. For `editvoice-edit`, use the same semantic model, tokenizer, speaker model, and distribution, together with `model_assets/checkpoints/kimi-flow.pt` and `model_assets/checkpoints/kimi-bigvgan/`. See the [EditVoice code repository README](https://github.com/DHY02/EditVoice#readme) for complete commands and input formats.

## License

This repository contains files under Apache-2.0 and MIT; no single license applies to every file. See [LICENSE](LICENSE) for the file-by-file license mapping and `licenses/` for the license texts. Third-party weights retain their original authorship and licenses.