allan-valin/phoneticxeus-onnx

PhoneticXeus — ONNX int8

An adapted (converted and quantised) copy of changelinglab/PhoneticXeus by Shikhar Bharadwaj, Chin-Jou Li, Kwanghee Choi, Eunjung Yeo, William Chen, Shinji Watanabe and David R. Mortensen (An Empirical Recipe for Universal Phone Recognition). Licensed, like the original, under CC BY-NC-SA 4.0. Non-commercial use only.

Changes made: exported to ONNX (raw 16 kHz mono waveform → frame-level CTC logits), the frontend's whole-utterance waveform normalisation moved into the exported graph, weights dynamically quantised to int8 with onnxruntime. Export script: speech-to-ipa/eval/export_pxeus_onnx.py.

Files

• model.int8.onnx — input wav float32 [1, samples], output logits [1, frames, vocab].
• ipa_vocab.json — symbol → id (blank = 0). Greedy CTC: argmax per frame, merge repeats, drop <…> tokens.

Accuracy

On 760 VoxAngeles utterances (8 from each of 95 languages), narrow scoring: PER 49.4 / PFER 14.6 for this int8 copy vs 48.9 / 14.0 for the original (the fp32 ONNX export matches the original exactly).

Used by speech-to-ipa, a browser-only speech-to-IPA tool.