---
library_name: transformers
pipeline_tag: audio-to-audio
tags:
- audio-classification
- signal-processing
license: apache-2.0
---





# DashengTokenizer

<div align="center">


<a href="https://arxiv.org/abs/2602.23765"><img src="https://img.shields.io/badge/arXiv-2602.23765-b31b1b" alt="version"></a>
 <a href="https://huggingface.co/mispeech/dashengtokenizer"><img src="https://img.shields.io/badge/HuggingFace-ffcc66" alt="version"></a>
 <a href="https://arxiv.org/abs/2602.2602.23765"><img src="https://img.shields.io/badge/license-Apache-13333b" alt="version"></a>
<a href="https://huggingface.co/mispeech/dashengtokenizer/colab">
  <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab">
</a>

</div>

DashengTokenizer is a high-performance continious audio tokenizer designed for audio understanding and generation tasks.
Compared to previous works, our framework trains a **single linear layer** to enable audio generation for semantically strong encoders.

Achievements:

* **State-of-the-Art** Audio Understanding: DashengTokenizer consistently outperforms most previous self-supervised and supervised audio encoders.
* **High-Fidelity** Signal Reconstruction: Maintains exceptional signal integrity, ensuring that audio remains crisp and accurate after processing.
* Accelerated **Audio Generation** Training: Achieves optimal performance significantly faster than standard VAE models, reducing training time and costs.
* Superior **Speech Enhancement**: Provides a more robust encoding foundation for isolating and clarifying speech in noisy environments.


![Framework](./figures/framework.png)

## Usage

### Installation

```bash
uv pip install transformers torch torchaudio einops
```

### Basic Usage

```python
import torch
import torchaudio
from transformers import AutoModel

# Load the model
model = AutoModel.from_pretrained("mispeech/dashengtokenizer", trust_remote_code=True)
model.eval()

# Load audio file (only 16kHz supported!)
audio, sr = torchaudio.load("path/to/audio.wav")

# Optional: Create attention mask for variable-length inputs
# attention_mask = torch.ones(audio.shape[0], audio.shape[1])  # All ones for full audio
# attention_mask[0, 8000:] = 0  # Example: mask second half of first sample

# Method 1: End-to-end processing (encode + decode)
with torch.no_grad(), torch.autocast(device_type='cuda'):
    outputs = model(audio)  # Optionally pass attention_mask=attention_mask
    reconstructed_audio = outputs["audio"]
    embeddings = outputs['embeddings']

# Method 2: Separate encoding and decoding
with torch.no_grad(), torch.autocast(device_type='cuda'):
    # Encode audio to embeddings
    embeddings = model.encode(audio)  # Optionally pass attention_mask=attention_mask

    # Decode embeddings back to audio
    reconstructed_audio = model.decode(embeddings)

# Save reconstructed audio
torchaudio.save("reconstructed_audio.wav", reconstructed_audio, sr)
```


## Use Cases

### 1. Audio Encoding
```python
embeddings = model.encode(audio)
reconstructed = model.decode(embeddings)
```

### 2. Feature Extraction
```python
# Extract rich audio features for downstream tasks
features = model.encode(audio)
# Use features for classification, clustering, etc.
```


## Limitations

- Optimized for 16kHz mono audio

## Results

![Audio Generation Results](./figures/audio_generation_results.png)
![Audio Understanding Results](./figures/audio_understanding_results.png)

## Citation

If you use DashengTokenizer in your research, please cite:

```bibtex
@misc{dinkel_dashengtokenizer_2026,
  title={DashengTokenizer: One layer is enough for unified audio understanding and generation},
  author={MiLM Plus, Xiaomi},
  year={2026},
  url={https://huggingface.co/mispeech/dashengtokenizer}
}
```

## License

Apache 2.0 License