codebucketsolutions/dicow-v3.3-onnx

DiCoW v3.3, ONNX

ONNX export of BUT-FIT/DiCoW_v3_3 (revision c34b64d9a9c5148c65fd355bb188d60343a6b44f) for bolo, which runs it with ONNX Runtime and NumPy, without PyTorch. DiCoW (Diarization-Conditioned Whisper) is Whisper large-v3-turbo conditioned on per-frame speaker activity (silence / target / non-target / overlap masks), so a recording is transcribed once per diarized speaker, overlapping speech included. The model is by the Speech@FIT group of Brno University of Technology and is licensed CC-BY-4.0; this conversion keeps that licence. Changes: the weights are exported unchanged to two graphs (encoder with FDDT conditioning and the decoder's cross-attention projections; decoder with a key/value cache); the *.int8.onnx graphs are dynamically quantised (int8 weights, per output channel; the encoder's feed-forward output matrices stay fp32). The CTC head is not included (the published generation settings do not use it). On 24 AMI excerpts the int8 graphs scored the same cpWER as the PyTorch model (26.0 %); details in bolo's docs/models.md, "DiCoW without PyTorch".

Exported with server/tools/export_dicow_onnx.py from the bolo repository; checksums in dicow_onnx.json.

If you use the model, please cite the papers its model card lists:

• A. Polok et al., "DiCoW: Diarization-conditioned Whisper for target speaker automatic speech recognition", Computer Speech & Language 95, 2026, doi:10.1016/j.csl.2025.101841.
• A. Polok et al., "Target Speaker ASR with Whisper", ICASSP 2025, doi:10.1109/ICASSP49660.2025.10887683.
• A. Polok, D. Klement, S. Cornell, M. Wiesner, J. Černocký, S. Khudanpur, L. Burget, "SE-DiCoW: Self-Enrolled Diarization-Conditioned Whisper", arXiv:2601.19194, 2026.