An ONNX export of facebook/wav2vec2-lv-60-espeak-cv-ft (Apache-2.0) — a wav2vec 2.0 CTC model fine-tuned to emit espeak IPA phonemes — for use as the acoustic model in sadda's phone-level forced aligner (sadda.align).

This is a format conversion, not a retrained model. The weights are Meta's, unchanged — only exported to ONNX (torch.onnx, opset 17) and cast to fp16. Reproducible recipe: tools/align/export_onnx.py.

Files

model.onnx — fp16, single file. Input: mono 16 kHz waveform (1, N), zero-mean/unit-variance normalized; output: (1, T, 392) CTC logits at 50 fps.
vocab.json — the 392-token espeak-IPA vocabulary (CTC blank = <pad> = 0).

Fidelity

The fp32 ONNX matches the PyTorch model to max|Δlogit| = 8.4e-4; the hosted fp16 produces identical word/phone boundaries to fp32 on test audio.

Usage

python

import sadda
model = sadda.align.Wav2Vec2EspeakModel.from_pretrained("sadda-speech/wav2vec2-espeak-ctc")
alignment = sadda.align.align(audio, 16000, transcript, model=model)

License

Apache-2.0, inherited from the source model; Meta's copyright/attribution retained.

Citation

Xu, Q., Baevski, A. & Auli, M. (2022). Simple and Effective Zero-shot Cross-lingual Phoneme Recognition. Interspeech 2022, 2113–2117. doi:10.21437/Interspeech.2022-60. https://arxiv.org/abs/2109.11680
Baevski, A., Zhou, Y., Mohamed, A. & Auli, M. (2020). wav2vec 2.0. NeurIPS 33, 12449–12460. https://arxiv.org/abs/2006.11477