MultIPA — ONNX int8

A converted and quantised copy of ctaguchi/wav2vec2-large-xlsr-japlmthufielta-ipa1000-ns (MultIPA) by Chihiro Taguchi, Yusuke Sakai, Parisa Haghani and David Chiang, Universal Automatic Phonetic Transcription into the International Phonetic Alphabet (Interspeech 2023). MIT licence, as granted by the author in ctaguchi/multipa (2026-10-19).

Changes made: exported to ONNX (raw 16 kHz mono waveform → frame-level CTC logits), the feature extractor's zero-mean/unit-variance normalisation moved into the graph, weights dynamically quantised to int8 with onnxruntime. Export script: speech-to-ipa/eval/export_multipa_onnx.py.

Files

model.int8.onnx — input wav float32 [1, samples], output logits [1, frames, vocab].
vocab.json — {"blank": 310, "vocab": {symbol: id}}; [PAD] is the CTC blank, | the word separator. Greedy CTC: argmax per frame, merge repeats, drop blank.

Accuracy

On 760 VoxAngeles utterances (8 from each of 95 languages), narrow scoring: PER 61.3 / PFER 15.1 for this int8 copy vs 61.1 / 14.6 for the original.

Used by speech-to-ipa.

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for allan-valin/multipa-onnx

Base model

ctaguchi/wav2vec2-large-xlsr-japlmthufielta-ipa1000-ns

Quantized

(1)

this model