PaperlayerView exact revision on arXiv

Source-linked AI summary

Group Relative Policy Optimisation Improves Multilingual Speech Recognition in Low-Resource Languages

Tim Wenjie Hu, Evelyn Nafula Ouma, Benjamin Akera, John A. Quinn

arXiv:2610.30663v1cs.AIcs.CL

TL;DR

Autoregressive ASR trained with teacher forcing can struggle to recover from its own prediction errors at inference. This paper applies GRPO to Whisper-large-v3 across 51 African languages, finding the strongest benefits in low-resource and noisy settings and reporting the best published average WER on AfriVox-v2 and SimbaBench.

  • Problem

    Teacher-forced autoregressive ASR trains on preceding tokens being correct, leaving models at risk of exposure bias when they must recover from prediction errors at inference.

    Source: §1 · p. 1

  • Method

    The paper applies GRPO to Whisper-large-v3 across 51 African languages, sampling and scoring complete transcript rollouts and training the decoder.

    Source: §1 · p. 1, §3 · p. 2

  • Results

    SunflowerASR achieves the best published average WER on AfriVox-v2 and SimbaBench, with gains most pronounced for noisy audio and languages with scarce training data.

    Source: §1 · p. 1

  • Takeaways & Limitations

    The reported findings support GRPO as a post-training method for autoregressive ASR, with benefits increasing in low-resource and challenging-audio settings.

    Source: §5 · p. 4

Abstract

from arXiv · show

Autoregressive encoder-decoder speech recognition models trained with teacher forcing can suffer from exposure bias, limiting their ability to recover from their own prediction errors at inference time. Recent work has shown that reinforcement learning with Group Relative Policy Optimisation (GRPO) improves speech recognition in major languages with LLM-based models. We extend this line of work to smaller, non-LLM encoder-decoder ASR models, using Whisper-large-v3 as a case study across 51 African languages. The benefits of this approach are most pronounced when training data is scarce or when input audio is noisy or degraded. Our resulting model, SunflowerASR, achieves the best published macro-averaged WER on the AfriVox-v2 (0.227) and SimbaBench (0.372) benchmarks, and is publicly available.

1. INTRODUCTION

The introduction motivates GRPO as a way to address exposure bias and asks whether its benefits extend to smaller non-LLM models and low-resource languages. The authors study Whisper-large-v3 across 51 African languages and report stronger gains on noisy audio and data-scarce languages.

1. INTRODUCTION

  • SFT trains on the assumption that every preceding token is correct, leaving autoregressive models at risk of exposure bias during inference.The model receives no training signal for recovering from its own preceding-token mistakes.Source: §1 · p. 1
  • The paper tests whether GRPO benefits extend from LLM-based ASR to smaller non-LLM encoder-decoder models and languages with scarce training data.Prior work cited in the introduction reports GRPO benefits with LLM-based speech recognition models.Source: §1 · p. 1
  • SunflowerASR achieves the best published average WER on AfriVox-v2 and SimbaBench, though per-language rankings vary.The study applies GRPO to Whisper-large-v3 across 51 African languages.Source: §1 · p. 1
  • GRPO gains are most pronounced for noisy test audio and languages with scarce training data.The case study uses Whisper-large-v3, a 1.55B-parameter model, to transcribe 51 African languages.Source: §1 · p. 1

2. DATASETS AND BASE MODEL

The training corpus combines public speech datasets across 51 African languages, while the accompanying figure compares CER before and after GRPO under increasing background noise.

2. DATASETS AND BASE MODEL

  • After filtering, the training corpus contains 3.82 million clips totaling 7,412 hours of speech from 110 dataset subsets across 51 African languages.It averages 2.2 independent sources per language; suspected noisy labels are filtered using an intermediate ASR checkpoint.Source: §2 · p. 1
  • The corpus draws on Common Voice, African Voices, Waxal, and other publicly available datasets.Full source and language-level details are available in the Hugging Face release.Source: §2 · p. 1
  • The figure compares Whisper-large-v3 CER before supervised fine-tuning only and after GRPO across 51 African languages; improvement grows with background noise.The figure caption identifies the comparison and noise trend but supplies no CER values.Source: Figure 1 · p. 1

3. REINFORCEMENT LEARNING WITH GRPO

GRPO samples and scores complete transcripts, using the rollout group’s mean reward as a baseline. The approach selects clips with headroom and trains the decoder using multilingual batches.

3. REINFORCEMENT LEARNING WITH GRPO

  • For each audio clip, GRPO samples a group of complete transcripts, scores each rollout, and ranks them against the group’s mean reward.The objective optimizes a sequence-level metric, unlike SFT’s per-token likelihood of one reference transcript.Source: §3 · p. 2
  • Probability mass shifts toward positive-advantage rollouts under a token-level normalized objective with a KL penalty of β = 0.01 against the initial policy.Source: §3 · p. 2
  • A clip whose rollouts all receive the same reward has zero advantage for every member and contributes no gradient, even if the clip is difficult.The method can learn from clips where a good transcript is possible but not reliably sampled.Source: §3 · p. 2
  • The method selects the 200 clips with the largest headroom from each training subset, totaling 15,285 clips.Headroom is the gap between a typical sample’s CER and the best sample’s CER in the group; a large gap indicates a better transcript is already in the sampling distribution.Source: §3 · p. 2
  • The authors compare headroom-based selection with randomly drawing the same number of clips per subset in an ablation.The comparison is described as Section 4.1’s ablation.Source: §3 · p. 2
  • GRPO training updates only the decoder’s 906 million parameters, using 10 rollouts per example and batches of 64 distinct clips across languages.The Whisper encoder is frozen; the batch is intended to average updates across languages rather than favor whichever language appears in a small batch.Source: §3 · p. 2

4. RESULTS

SunflowerASR outperforms published alternatives across three benchmarks, with larger gains under degraded audio and for languages with less supervised data. Deployment and ablation results also show that GRPO’s benefits depend on training choices and extend to difficult real-world recordings.

4. RESULTS

  • GRPO consistently improves CER over the SFT checkpoint, with larger improvements as audio degradation becomes more severe.The comparison uses audio with three noise levels and undegraded audio as a control.Source: §4 · p. 2
  • SunflowerASR has superior accuracy to all published alternatives on Sunbird, AfriVox-v2, and SimbaBench.The supplied passages do not provide benchmark-level result values.Source: §4 · p. 2, Figure 2 · p. 3, Table 1 · p. 3

4.1. Ablations

  • One extra supervised epoch on GRPO’s same 15,285 clips regresses both WER and CER, although those clips help under the reinforcement objective.The ablation continues supervised fine-tuning from the same checkpoint using exactly the clips GRPO received.Source: §4.1 · p. 3, Table 4 · p. 4
  • Random selection reaches WER 0.4220 and CER 0.1714, improving over SFT but recovering only about two-thirds of the CER gain (−3.9% versus −5.8% relative).This ablation replaces headroom ranking with uniform sampling while holding the rest fixed.Source: §4.1 · pp. 3–4

4. RESULTS

  • On 40 languages, CER improvement correlates with supervised-data hours at r = −0.41 on log10 hours, with 35 languages improving.Fig. 3 plots per-language CER improvement against the hours of supervised audio contributed to training.Source: §4 · pp. 2–3, Figure 3 · p. 4
  • On 60 deployment-evaluation clips, GRPO reduces aggregate WER by 6.2% relative and CER by 9.7% relative.The set covers Ugandan English and Luganda in noisy, spontaneous conversational settings.Source: §4 · p. 3, Table 3 · p. 4

4.1. Ablations

  • Reducing the batch from 64 to 16 clips yields WER 0.4243 and CER 0.1689, with a clearer regression on WER (−0.8% versus −1.8% relative).The variants use the same selected clips and G = 10 rollouts.Source: §4.1 · p. 4, Table 4 · p. 4

4.2. Characterisation of improvements

  • Under acoustic stress, GRPO repairs repetition loops; on remaining non-looping rows, word substitutions are flat at +0.5% while character substitutions fall by 3.0%.The supervised model also sometimes replaces words with fragments or unrelated compounds where GRPO produces near-miss spellings.Source: §4.2 · p. 4, Table 5 · p. 4, Table 6 · p. 4

5. CONCLUSION

GRPO transfers to a smaller non-LLM encoder-decoder ASR model in multilingual, low-resource, noisy settings. Its benefits are greater for languages with fewer training resources and challenging audio, while example selection and batch sizing help attain accuracy improvements.

  • GRPO transfers from LLM-decoder systems to a smaller non-LLM encoder-decoder model in multilingual, low-resource, and noisy settings.Source: §5 · p. 4
  • Benefits increase for languages with fewer data resources, including those with tens of hours of audio, and for challenging test audio.Source: §5 · p. 4
  • Selecting training examples with headroom and using appropriate batch sizes help attain accuracy improvements.Source: §5 · p. 4
  • The findings suggest GRPO may be a widely generalisable post-training method for autoregressive ASR models.Source: §5 · p. 4
Loading 2610.30663v1…