Group scoring helps speech models avoid teacher errors
Pseudo-Labels as Rewards, Not Targets: Group-Relative Policy Optimization for Semi-Supervised ASR · arXiv:2610.32572
In plain words
Speech recognizers can struggle with new kinds of audio, even when someone can provide audio but no transcripts. A larger teacher model can make rough transcripts, but a student trained to copy every word may also learn the teacher’s mistakes.
This paper tests a different approach: use the rough transcripts to score several possible student transcripts, rather than teach every word as correct. On one banking test set, this approach had lower word error rate than standard training on the same teacher’s transcripts.
The research
When audio from a new domain has no transcripts, a student recognizer can be trained on a teacher’s rough transcripts. But if training treats every word as correct, how can the student adapt without simply learning the teacher’s mistakes?
GRPO adapts a speech recognizer by scoring sampled transcripts against machine-made references, reducing how strongly it learns their errors.
The authors compared GRPO with SFT using the same Whisper Tiny pseudo-labels for banking audio. On the DefinedAI banking test set, evaluated against gold transcripts, GRPO had 14.06% word error rate versus 32.11% for SFT.
This result shows lower test error in that condition, while the error-copying checks offer evidence about how the methods differ.
Gold transcripts helped identify mistaken pseudo-label words. Direct GRPO repeated fewer of these words than SFT, and its sampled alternatives contained them less often.
This pattern fits the proposed mechanism: a mistaken word that candidates rarely produce has less influence on their relative scores.
The banking comparison changes both the training procedure and learning rate, and the paper does not say whether both students started from the same checkpoint.
The findings are specific to the tested data and training conditions; GRPO also required an initial supervised training stage for the four language settings where the English-trained student could not produce the target scripts.
What the evidence shows
- The risk in rough transcripts
Automatic speech recognition (ASR) systems can perform poorly on audio from a new domain, where transcripts may be unavailable. A common workaround is for a teacher system to transcribe the audio, then train a student on those machine-generated pseudo-labels.
But supervised fine-tuning (SFT) treats every pseudo-label word as correct, so teacher mistakes become student training targets.
- Score alternatives as a group
The paper proposes Group Relative Policy Optimization (GRPO), a method for adapting an ASR student using pseudo-labels as scoring references. For each utterance, the student produces four candidate transcripts.
GRPO scores them against the pseudo-label and compares their scores within that group, rather than training the student to predict each pseudo-label word. When every candidate makes the same omission, the shared penalty cancels out of their relative scores.
- Banking test and error checks
On the DefinedAI banking test set, GRPO trained with Whisper Tiny pseudo-labels had 14.06% word error rate, compared with 32.11% for SFT trained on those same pseudo-labels. The authors also checked how often students repeated pseudo-label mistakes: direct GRPO copied fewer erroneous words into its final transcripts than SFT, and those errors appeared less often in its sampled candidates. These observations support the proposed account of why group scoring can reduce error copying.
- Correct transcripts change the result
A banking control with gold training transcripts yielded a different ordering: SFT followed by GRPO had the lowest word error rate, while direct GRPO had the highest of the three tested approaches.
This comparison supports the authors’ view that GRPO’s advantage over SFT is linked to noisy references, rather than being a general advantage when the training transcripts are correct.
- Where the results extend
The reported pattern also appears in healthcare and insurance tests using Whisper Large-V3 pseudo-labels. In four low-resource language tests, character error rate fell when GRPO followed SFT.
Direct GRPO was not tested for those languages because the English-trained student could not initially produce their writing systems; SFT first was needed to establish the target script.
Deep
- Original paper authors
- Shashi Kumar, Yanis Labrak, Sergio Burdisso, Esaú Villatoro-Tello, Petr Motlicek
- Submitted to arXiv
- License
- CC BY 4.0