| --- |
| license: apache-2.0 |
| pipeline_tag: feature-extraction |
| tags: |
| - m5-tts |
| - audio-tokenizer |
| - audio-codec |
| - pytorch |
| --- |
| |
| # M5-TTS-Audio-Tokenizer |
|
|
| English | [简体中文](README_zh.md) |
|
|
| [GitHub](https://github.com/ASLP-lab/M5-TTS) · [Demo](https://aslp-lab.github.io/M5-TTS-Demo-Page/) · [Space](https://huggingface.co/spaces/ASLP-lab/M5-TTS-Space) · [Models](https://huggingface.co/collections/ASLP-lab/m5-tts) |
|
|
| This repository provides the **Qwen 12Hz audio codec used by M5-TTS**, encoding reference recordings into discrete audio tokens and decoding generated tokens into speech. It is **not a standalone text-to-speech model** and is distinct from the text tokenizer bundled with M5 checkpoints. |
|
|
| | Property | Value | |
| | --- | --- | |
| | Audio sample rate | 24 kHz | |
| | Codec frame rate | 12.5 frames/s | |
| | Codebooks used by M5 | 16 | |
| | Codes layout through the M5 adapter | `(batch, frames, 16)` for a padded batch | |
|
|
| The [VoiceClone](https://huggingface.co/ASLP-lab/M5-TTS-50M-VoiceClone) and [CustomVoice](https://huggingface.co/ASLP-lab/M5-TTS-50M-CustomVoice) inference scripts download this codec automatically. You normally do not need to load it separately. |
|
|
| ## Quickstart |
|
|
| For standalone audio encoding and reconstruction, use Python 3.10 and the M5 repository: |
|
|
| ```bash |
| git clone https://github.com/ASLP-lab/M5-TTS.git |
| cd M5-TTS |
| conda create -n m5tts python=3.10 -y |
| conda activate m5tts |
| pip install -r requirements.txt |
| ``` |
|
|
| Run from the repository root; this CPU example does not require FlashAttention: |
|
|
| ```python |
| import soundfile as sf |
| from m5.codec_qwen12hz import Qwen3TTSTokenizer |
| |
| codec = Qwen3TTSTokenizer.from_pretrained( |
| "ASLP-lab/M5-TTS-Audio-Tokenizer", |
| device_map="cpu", |
| ) |
| |
| encoded = codec.encode("assets/aslp_default_voice.wav") |
| waveforms, sample_rate = codec.decode(encoded) |
| sf.write("reconstructed.wav", waveforms[0], sample_rate) |
| ``` |
|
|
| `encode()` also accepts NumPy waveforms with an explicit source sample rate, for example `codec.encode(audio, sr=24000)`. The adapter resamples input audio as needed; reconstruction is lossy. The TTS inference scripts apply their reference-audio preprocessing separately. |
|
|
| ## License and Attribution |
|
|
| Apache-2.0. This repository uses the [Qwen3-TTS 12Hz tokenizer](https://huggingface.co/Qwen/Qwen3-TTS-Tokenizer-12Hz). The M5 repository includes its minimal codec implementation with the [upstream license and modification notice](https://github.com/ASLP-lab/M5-TTS/tree/main/m5/codec_qwen12hz). No separate `qwen-tts` package installation is required. |
|
|
