CantoCaptions Cantonese ASR

• A fine-tuning of Qwen3-ASR on a subset of the CantoCaptions dataset

Details

This model is designed for high-quality written Cantonese transcription using the CantoCaptions standards, which mostly follow acceptable usages and variants documented in the words.hk dictionary. Sentence final particles (SFPs) such as 啦 laa1 / 喇 laa3 / 嗱 laa4 are separated out by tone as described on the CantoCaptions website.

The current model was trained using LoRa rank=128 and trained over a single epoch on ~75h of training audio, with an additional 4h dev and 4h test audio derived from the same dataset. At the time of writing, the model is in an early stage of development and may be updated.

Evals

When accounting for character variants such as 啦/喇, 呀/啊, 咁/噉, and 畀/俾, the latest version of this model has a CER of 3.81% against the Mozilla Common Voice 13 Cantonese test set. It significantly outperforms Qwen on this metric even when converting all text to simplified Chinese for the evaluation.