tatsu020/mms-300m-fa-onnx

MMS-300M Forced Aligner — ONNX (fp16)

ONNX conversion of MahmoudAshraf/mms-300m-1130-forced-aligner for use with ONNX Runtime (e.g. onnxruntime-node with DirectML/CUDA).

Attribution / 帰属

• Original model: Meta AI — MMS-300M forced-alignment checkpoint (from torchaudio.pipelines.MMS_FA), part of the Massively Multilingual Speech (MMS) project.
• HF Transformers conversion: MahmoudAshraf/mms-300m-1130-forced-aligner (used with ctc-forced-aligner).

Changes made / 変更点

• Exported Wav2Vec2ForCTC to ONNX (opset 17, dynamic batch/samples axes, input: input_values [batch, samples] float32, output: logits [batch, frames, 31] float32).
• Converted weights to fp16 (keep_io_types=True, so I/O stays float32).
• No fine-tuning, no architecture changes.

Files

file	size	sha256
mms_300m_fa_fp16.onnx	631,499,926 bytes	9823aa1e22dfc91fe1b1185ead5db5cab4512ce604ed18459b17a9eb92a00cd2
vocab.json	—	token vocabulary + blank id + frame metadata

Usage notes

• Input: 16 kHz mono float32 waveform, zero-mean/unit-variance normalized per chunk.
• Frame stride: 320 samples (20 ms per output frame).
• CTC blank id: 0 (<blank>). Vocabulary is romanized (lowercase a–z + apostrophe).
• Intended for CTC forced alignment (e.g. lyrics-to-audio alignment), not ASR decoding.

License / ライセンス

CC-BY-NC-4.0 (inherited from the original Meta MMS checkpoint). Non-commercial use only. / 非商用利用に限ります。

This repository exists to distribute the ONNX conversion for the free, non-commercial karaoke practice app UtaLog.