Reinforcement Learning for Data-Efficient Code-Switched ASR

Researchers propose a practical reinforcement learning with verifiable rewards recipe for data-efficient adaptation of audio-language models to code-switched automatic speech recognition (ASR). The method uses group relative policy optimization, combining an error rate reward with a script fidelity reward that penalizes wrong writing systems.

• Training on only TTS code-switched speech using Qwen2-Audio across 10 language pairs.
• RLVR with 10% of the data matches LoRA supervised fine-tuning trained on the full dataset.
• The error rate reward eliminates translation errors while the script fidelity reward reduces script contamination.
• Gains transfer zero-shot to a human-recorded code-switching corpus.

The approach shows largest gains on typologically distant pairs and provides a reproducible testbed for adapting models to code-switched speech without requiring large labeled datasets.