Improved Gemini audio models for powerful voice experiences

Google DeepMind · 7 months ago

Google released an updated Gemini 2.5 Flash Native Audio model that improves voice agents' ability to handle complex workflows, follow instructions, and maintain natural conversations. The model achieved a 71.5% score on ComplexFuncBench Audio for multi-step function calling and reached a 90% adherence rate to developer instructions, up from 84%. The update enables live voice agents across Google products and introduces live speech-to-speech translation in Google Translate, supporting over 70 languages while preserving speaker characteristics like intonation and pacing.