GRPO treats ASR pseudo-labels as rewards rather than targets
Original title: Pseudo-Labels as Rewards, Not Targets: Group-Relative Policy Optimization for Semi-Supervised ASR
In banking audio, GRPO cut WER 29.3% to 56.2% relative to SFT and copied 27% of teacher errors.
The paper argues that supervised fine-tuning on pseudo-labels makes automatic speech recognition students copy teacher mistakes. Using the same labels as a reward for Group Relative Policy Optimization is presented as more robust to noise. In the banking domain, GRPO gave 29.3% to 56.2% relative word-error-rate gains over SFT on Whisper-generated labels. The authors say the advantage held across four Whisper sizes, four teacher families, three domains, and four languages. SFT reproduced almost 100% of teacher errors verbatim; GRPO passed through 27%.
Unconfirmed: WER and error-passthrough figures are the authors’ experimental claims.
- SFT treats every pseudo-label token as a training target.
- The reported edge spans four Whisper sizes and four languages.
- Erroneous pseudo-label words rarely appear in on-policy rollouts.
- SFT is said to collapse the reward variance GRPO uses.
From the original article
1 min read at arxiv.orgAdapting ASR to a domain with only untranscribed audio typically relies on pseudo-labeling, a strong teacher transcribes the audio and a student is fine-tuned on it. Because supervised fine-tuning (SFT) treats every pseudo-label token as a target, the student inherits the teacher's errors.
By Shashi Kumar · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.