yi-whisper-large-v3-ONNX

yi-whisper-large-v3 — ONNX

An ONNX build of ivrit-ai/yi-whisper-large-v3 for running Yiddish speech recognition in the browser with transformers.js.

No weights were retrained. This is a format conversion and quantisation, so that the model can be downloaded once by a web page and run locally on the user's own machine — audio never leaves the device. Used in production at yiddishtools.org.

Variants: q4f16 (GPU) and q4 (CPU).

Quality

On a held-out benchmark of six minutes of hard oral-history audio — an elderly speaker, dialect vowels, tape noise — this build transcribes at roughly 17% character error / 32% word error under the greedy decoding a browser performs. A good share of the remaining differences are spelling variants rather than mistakes. The output is a draft to correct, not a finished transcript.

Attribution

The underlying model is a Yiddish finetune of openai/whisper-large-v3 trained by ivrit-ai and released under Apache-2.0. All credit for the Yiddish capability belongs to them; this repository changes the file format.

Note on the translate task

Per the upstream model card, the translation task was not trained and is degraded. Use task: "transcribe" with language: "yi", and translate the resulting text separately.

Methodology

The conversion and evaluation methodology will be described in a forthcoming paper; this card will be updated with a citation when it is available.