dellusional/sepformer-wsj02mix-int8-onnx

SepFormer (wsj0-2mix) — int8 ONNX export

An int8-quantized ONNX export of SpeechBrain's sepformer-wsj02mix speech-separation model, packaged for plain-URL download by DontSpeak — 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 (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 for the original work — please cite SpeechBrain and the SepFormer paper (Subakan et al., 2021) if you use this model in research.