---
license: apache-2.0
language:
- en
tags:
- speech-separation
- sepformer
- onnx
- int8
- audio-to-audio
base_model: speechbrain/sepformer-wsj02mix
pipeline_tag: audio-to-audio
---

# SepFormer (wsj0-2mix) — int8 ONNX export

An int8-quantized ONNX export of SpeechBrain's
[`sepformer-wsj02mix`](https://huggingface.co/speechbrain/sepformer-wsj02mix)
speech-separation model, packaged for plain-URL download by
[DontSpeak](https://dontspeak.org) — the voice interface for coding agents.

DontSpeak uses it for the dictation **speaker-lock**: the "talk over a YouTube
video" case that frame-gating (diarization) can't solve. The model splits a
single-mic mixture into its constituent voices; the caller embeds each stream
with a speaker-ID model and keeps the one matching the enrolled user, so a
co-channel background voice is *removed* (not merely gated) before transcription.

## Files

| file | size | sha256 |
| --- | --- | --- |
| `sepformer_int8.onnx` | 29,927,349 B | `c28ef4168295b182fbf4b18b3c5743d649d39cbf7e0eee8b0e49a653f35bcb5e` |

## Model contract

- **Input**: mono PCM float32, **8 kHz** (the wsj0-2mix native rate), shape `[1, T]`
  with a **dynamic time axis** — resample 16 kHz capture to 8 kHz before the call.
- **Output**: **2 separated sources**, `[1, T, 2]`.
- **Runtime note**: because of the dynamic time axis, execution providers that
  recompile per input shape (e.g. CoreML) are pathologically slow — run it on the
  **CPU EP** (RTF ≈ 0.4 on Apple Silicon). Static-shape models are unaffected.

## Provenance

- Base model: [`speechbrain/sepformer-wsj02mix`](https://huggingface.co/speechbrain/sepformer-wsj02mix)
  (SpeechBrain, Apache-2.0), trained on wsj0-2mix.
- Export: traced to ONNX and int8-quantized (dynamic quantization) in June 2026 for the
  DontSpeak project. The one-off export script was not preserved; the artifact is pinned
  by the sha256 above and validated end-to-end by DontSpeak's speaker-lock tests.

## License

Apache-2.0, inherited from SpeechBrain and the base model. See
[SpeechBrain](https://github.com/speechbrain/speechbrain) for the original work — please
cite SpeechBrain and the SepFormer paper ([Subakan et al., 2021](https://arxiv.org/abs/2010.13154))
if you use this model in research.