This is true with natural language as well in LLMs, ASRs, and more. At Rev we found that adding data from other languages to primarily-English ASR models helpe

This is true with natural language as well in LLMs, ASRs, and more. At Rev we found that adding data from other languages to primarily-English ASR models helped reduce WER in the target language substantially (especially when accents were involved). We found similar benefits when training translation models.