Abstract
Speech Language Models (SLMs) that understand spoken language questions and commands support only a few high-resource languages, limiting access to modern technology for millions of speakers worldwide. This gap in language coverage stems from the scarcity of multilingual speech-language instruction-tuning datasets. To address this issue, we present MultiSpeechQA, a large-scale, synthetically generated and human-verified dataset comprising 9200 hours of more than 10.8 million spoken question-answer pairs in 23 typologically diverse languages, designed to improve the multilingual instruction-following capabilities of SLMs. Using MultiSpeechQA, we also introduce MultiSpeech-Bench, a multi-task benchmark to evaluate SLM performance across 23 languages. We compare the performance of a strong cascading system to three leading open-weight SLMs on MultiSpeech-Bench and find that the cascading system outperforms all existing open-weight SLMs. We then demonstrate the effectiveness of MultiSpeechQA by fine-tuning the best-performing open-weight SLM, Qwen 2.5-Omni, on our dataset, which substantially improves its performance and establishes new state-of-the-art results for open-weight models on our benchmark. Our findings show that high-quality synthetic datasets offer a scalable solution to improving the multilingual capabilities of SLMs, extending the benefits of natural spoken interactions to a wider range of languages.