File size: 8,219 Bytes

---
language:
- en
- zh
license: other
license_name: license-term-of-stabletoken
tags:
- speech tokenizer
pipeline_tag: audio-to-audio
---

# StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs (ICLR 2026)

**StableToken** is a noise-robust semantic speech tokenizer that performs discrete speech representation learning, achieving state-of-the-art stability in noisy environments.

📄 [Paper](https://huggingface.co/papers/2509.22220) | 💻 [GitHub](https://github.com/Tencent/StableToken)

For code and more detailed information, please refer to the corresponding [GitHub repository](https://github.com/Tencent/StableToken).

## Model Details

| Attribute | Value |
|:----------|:------|
| Frame Rate | 25 Hz |
| Codebook Size | 8,192 |
| BPS (Bits Per Second) | 325 |

## Quick Start

To use StableToken, please clone the official repository and install dependencies.

### Installation

```bash
git clone --recursive https://github.com/Tencent/StableToken.git
cd StableToken && pip install -r requirements.txt
```

### Inference

```python
import os
from huggingface_hub import snapshot_download
from transformers import WhisperFeatureExtractor
from src.model.modeling_whisper import WhisperLFQEncoder
from src.utils.flow_inference import AudioDecoder
from src.utils.utils import extract_speech_token, speech_token_to_wav

# 1. Download & Load Models
model_dir = snapshot_download("tencent/StableToken")

# Load Tokenizer
tokenizer = WhisperLFQEncoder.from_pretrained(os.path.join(model_dir, "tokenizer")).eval().cuda()
feature_extractor = WhisperFeatureExtractor.from_pretrained(os.path.join(model_dir, "tokenizer"))

# Load Decoder
decoder = AudioDecoder(
    config_path=os.path.join(model_dir, "decoder", "config.yaml"),
    flow_ckpt_path=os.path.join(model_dir, "decoder", "flow.pt"),
    hift_ckpt_path=os.path.join(model_dir, "decoder", "hift.pt"),
    device="cuda"
)

# 2. Tokenize
tokens = extract_speech_token(tokenizer, feature_extractor, ["/path/to/audio.wav"], device="cuda")[0]

# 3. Reconstruct
tts_speech, sampling_rate = speech_token_to_wav(decoder, tokens)
```

## Performance

StableToken achieves **60% lower UED** (Unit Edit Distance) than best existing supervised semantic tokenizers.

### Noise Robustness (UED ↓)

| Model | Frame Rate | Codebook Size | UED (%, ↓) |
|:---|:---:|:---:|:---:|
| [GLM-4-Voice-Tokenizer](https://github.com/zai-org/GLM-4-Voice) | 12.5Hz | 16,384 | 31.10 |
| [S3 Tokenizer](https://github.com/FunAudioLLM/CosyVoice) | 25Hz | 4,096 | 26.17 |
| [CosyVoice2](https://github.com/FunAudioLLM/CosyVoice) | 25Hz | 6,561 | 38.66 |
| **StableToken** | 25Hz | 8,192 | **10.17** 🏆 |

### Reconstruction Quality

Measurements on LibriSpeech (LS) and SEED benchmarks.

| Model | Frame<br>Rate | BPS | WER (↓)<br>LS-clean | WER (↓)<br>LS-other | WER (↓)<br>SEED-en | WER (↓)<br>SEED-zh | MOS (↑)<br>LS-clean | MOS (↑)<br>LS-other | MOS (↑)<br>SEED-en | MOS (↑)<br>SEED-zh |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
| [GLM-4-Voice-Tokenizer](https://github.com/zai-org/GLM-4-Voice) | 12.5Hz | 175 | 4.04 | 9.33 | 3.54 | 3.23 | 4.07 | **3.99** | **4.16** | 4.10 |
| [S3 Tokenizer](https://github.com/FunAudioLLM/CosyVoice) | 25Hz | 300 | 5.78 | 13.38 | 5.91 | 4.26 | 3.40 | 3.31 | 3.40 | 3.31 |
| [CosyVoice2](https://github.com/FunAudioLLM/CosyVoice) | 25Hz | 325 | 4.25 | 9.68 | 4.34 | 2.75 | 3.36 | 3.25 | 3.31 | 3.58 |
| **StableToken** | 25Hz | 325 | **3.84** | **7.99** | **3.44** | **2.62** | **4.09** | 3.83 | 4.01 | **4.18** |

## Citation

```bibtex
@article{song2025stabletoken,
  title={StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs},
  author={Song, Yuhan and Zhang, Linhao and Wu, Chuhan bitwise voting mechanism to form a single, stable token sequence. StableToken sets a new state-of-the-art in token stability, drastically reducing Unit Edit Distance (UED) under diverse noise conditions. This foundational stability translates directly to downstream benefits, significantly improving the robustness of SpeechLLMs on a variety of tasks. Our code and model are publicly available at this https URL .

# Current model card

The README of the model repository currently looks like this:

## Metadata
```yaml
language:
- en
- zh
license: other
license_name: license-term-of-stabletoken
tags:
- speech tokenizer
```

## Content
# StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs (ICLR 2026)

**StableToken** is a noise-robust semantic speech tokenizer that performs discrete speech representation learning, achieving state-of-the-art stability in noisy environments.

📄 [Paper](https://arxiv.org/abs/2509.22220) | 💻 [GitHub](https://github.com/Tencent/StableToken)

For code and more detailed information, please refer to the corresponding [GitHub repository](https://github.com/Tencent/StableToken).

## Model Details

| Attribute | Value |
|:----------|:------|
| Frame Rate | 25 Hz |
| Codebook Size | 8,192 |
| BPS (Bits Per Second) | 325 |

## Quick Start

To use StableToken, please clone the official repository and install dependencies.

### Installation

```bash
git clone --recursive https://github.com/Tencent/StableToken.git
cd StableToken && pip install -r requirements.txt
```

### Inference

```python
import os
from huggingface_hub import snapshot_download
from transformers import WhisperFeatureExtractor
from src.model.modeling_whisper import WhisperLFQEncoder
from src.utils.flow_inference import AudioDecoder
from src.utils.utils import extract_speech_token, speech_token_to_wav

# 1. Download & Load Models
model_dir = snapshot_download("tencent/StableToken")

# Load Tokenizer
tokenizer = WhisperLFQEncoder.from_pretrained(os.path.join(model_dir, "tokenizer")).eval().cuda()
feature_extractor = WhisperFeatureExtractor.from_pretrained(os.path.join(model_dir, "tokenizer"))

# Load Decoder
decoder = AudioDecoder(
    config_path=os.path.join(model_dir, "decoder", "config.yaml"),
    flow_ckpt_path=os.path.join(model_dir, "decoder", "flow.pt"),
    hift_ckpt_path=os.path.join(model_dir, "decoder", "hift.pt"),
    device="cuda"
)

# 2. Tokenize
tokens = extract_speech_token(tokenizer, feature_extractor, ["/path/to/audio.wav"], device="cuda")[0]

# 3. Reconstruct
tts_speech, sampling_rate = speech_token_to_wav(decoder, tokens)
```

## Performance

StableToken achieves **60% lower UED** (Unit Edit Distance) than best existing supervised semantic tokenizers.

### Noise Robustness (UED ↓)

| Model | Frame Rate | Codebook Size | UED (%, ↓) |
|:---|:---:|:---:|:---:|
| [GLM-4-Voice-Tokenizer](https://github.com/zai-org/GLM-4-Voice) | 12.5Hz | 16,384 | 31.10 |
| [S3 Tokenizer](https://github.com/FunAudioLLM/CosyVoice) | 25Hz | 4,096 | 26.17 |
| [CosyVoice2](https://github.com/FunAudioLLM/CosyVoice) | 25Hz | 6,561 | 38.66 |
| **StableToken** | 25Hz | 8,192 | **10.17** 🏆 |

### Reconstruction Quality

Measurements on LibriSpeech (LS) and SEED benchmarks.

| Model | Frame<br>Rate | BPS | WER (↓)<br>LS-clean | WER (↓)<br>LS-other | WER (↓)<br>SEED-en | WER (↓)<br>SEED-zh | MOS (↑)<br>LS-clean | MOS (↑)<br>LS-other | MOS (↑)<br>SEED-en | MOS (↑)<br>SEED-zh |
|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
| [GLM-4-Voice-Tokenizer](https://github.com/zai-org/GLM-4-Voice) | 12.5Hz | 175 | 4.04 | 9.33 | 3.54 | 3.23 | 4.07 | **3.99** | **4.16** | 4.10 |
| [S3 Tokenizer](https://github.com/FunAudioLLM/CosyVoice) | 25Hz | 300 | 5.78 | 13.38 | 5.91 | 4.26 | 3.40 | 3.31 | 3.40 | 3.31 |
| [CosyVoice2](https://github.com/FunAudioLLM/CosyVoice) | 25Hz | 325 | 4.25 | 9.68 | 4.34 | 2.75 | 3.36 | 3.25 | 3.31 | 3.58 |
| **StableToken** | 25Hz | 325 | **3.84** | **7.99** | **3.44** | **2.62** | **4.09** | 3.83 | 4.01 | **4.18** |

## Citation

```bibtex
@article{song2025stabletoken,
  title={StableToken: A Noise-Robust Semantic Speech Tokenizer for Resilient SpeechLLMs},
  author={Song, Yuhan and Zhang, Linhao and Wu, Chuhan and Liu, Aiwei and Jia, Wei and Wang, Houfeng and Zhou, Xiao},
  journal={arXiv preprint arXiv:2509.22220},
  year={2025}
}
```

## License

This project is licensed under the [License Term of StableToken](LICENSE).
```