Vivek2259/mms-fa-mm8

MMS forced-alignment model, ONNX (MatMul int8)

Browser build of Meta's MMS forced-alignment model (torchaudio.pipelines.MMS_FA, wav2vec2, 300M params), used by the Caption Re-timer web app to time-align a known transcript to audio with CTC forced alignment.

• mms_fa_mm8.onnx (~356 MB): input audio float32 [1, samples] at 16 kHz → output logprobs [1, frames, 28], one frame per 20 ms. Input normalisation and log_softmax are inside the graph. Only the transformer MatMul weights are quantized to int8 (per channel); full int8 quantization caused multi-second misses in testing.
• Vocabulary (id 0 = CTC blank): - a i e n o u t s r m k l d g h y b p w c v j z f ' q x

Built with export_model.py (torch/torchaudio 2.5.1, opset 17).

Licence

The MMS weights are released by Meta under CC-BY-NC 4.0 (non-commercial use only). This is a converted copy of those weights under the same licence.