Voxtral Mini 4B Realtime Arabic released

Mistral AI's Voxtral Mini 4B Realtime Arabic produces text as speech arrives, supporting live captions and voice interfaces for Arabic dialects and Modern Standard Arabic. It is adapted from the earlier realtime model, with training focused on conversations where speakers switch languages. The model card describes configurable transcription delay, allowing applications to choose how long the model waits before producing text.

At a configured delay of 480 milliseconds, Mistral reports an average character error rate of 8.82% across seven Arabic benchmarks. This measures transcription mistakes at the character level; the result uses the team's Arabic text-normalization procedure, so scores computed with different spelling normalization may differ. Benchmark conditions.