Mistral Releases Voxtral: Open Source Speech Understanding Models (3B and 24B)