Group Relative Policy Optimisation Improves Multilingual Speech Recognition in Low-Resource Languages
3 judges. Scored by all three judges. Only the day’s top stories get the claim check. How far stories are checked
This paper applies Group Relative Policy Optimisation, a reinforcement learning method, to fine-tune Whisper-large-v3 for speech recognition across 51 African languages, aiming to reduce errors from exposure bias.
Score
0.60total under rubric v2.0
- Significance0.57
- Reader interest0.69
- Audience breadth0.43
- Practical relevance0.66
- Substantiated0.84
- Technical depth0.90
- New information0.90
- Counts against: Promotional0.10
- Counts against: Overclaim0.45
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| Significance | 0.56 | 0.50 | 0.66 | 0.57 | 0.16 | +0.25 | +0.144 |
| Reader interest | 0.64 | 0.60 | 0.82 | 0.69 | 0.22 | +0.20 | +0.137 |
| Audience breadth | 0.41 | 0.40 | 0.48 | 0.43 | 0.08 | +0.10 | +0.043 |
| Practical relevance | 0.74 | 0.55 | 0.69 | 0.66 | 0.19 | +0.15 | +0.099 |
| Substantiated | 0.96 | 0.70 | 0.86 | 0.84 | 0.26 | +0.10 | +0.084 |
| Technical depth | 0.97 | 0.80 | 0.93 | 0.90 | 0.17 | +0.10 | +0.090 |
| New information | 0.92 | 0.80 | 0.98 | 0.90 | 0.18 | +0.10 | +0.090 |
| Promotional | 0.06 | 0.20 | 0.04 | 0.10 | 0.16 | −0.15 | −0.015 |
| Overclaim | 0.43 | 0.40 | 0.52 | 0.45 | 0.12 | −0.15 | −0.068 |
| Sum of contributions | +0.605 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.605 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN —
Recorded for context only, checked 24 Oct 2026, 08:54 UTC. It played no part in the ranking.