---
license: cc-by-4.0
datasets:
- openslr/librispeech_asr
language:
- en
pipeline_tag: audio-to-audio
---

# SSLZip

## Usage

```py
import onnxruntime as ort
from transformers import HubertModel
import torch

# Load the upstream HuBERT model.
upstream = HubertModel.from_pretrained("facebook/hubert-base-ls960")
upstream.eval()

# Load the autoencoder model.
postprocessor = ort.InferenceSession("sslzip_256.onnx")
node_name = postprocessor.get_inputs()[0].name

# Prepare an input waveform (assuming 16kHz audio).
x = torch.randn(1, 16000)

# Extract the latent representation for downstream tasks.
with torch.inference_mode():
    h = upstream(x, output_hidden_states=True).hidden_states[-1]
    z = postprocessor.run(None, {node_name: h.cpu().numpy()})[0]

# Use z as you like.
print(z.shape)
```

## License

The pretrained model was developed using the LibriSpeech corpus and is distributed under the same license (CC BY 4.0).  
Please include credit to Nagoya Institue of Technology and Techno-Speech, Inc. when using this model.

## Citation

```bibtex
@InProceedings{yoshimura2025sslzip,
  author = {Takenori Yoshimura and Shinji Takaki and Kazuhiro Nakamura and Keiichiro Oura and Takato Fujimoto and Kei Hashimoto and Yoshihiko Nankaku and Keiichi Tokuda},
  title = {{SSLZip}: Simple autoencoding for enhancing self-supervised speech representations in speech generation},
  booktitle = {13th ISCA Speech Synthesis Workshop (SSW 2025)},
  pages = {xxx--xxx},
  year = {2025},
}
```