AudioPaLM is a multimodal language model that combines text-based and speech-based models for speech understanding and generation, with applications including speech recognition and speech-to-speech translation. It outperforms existing systems and can perform zero-shot speech-to-text translation for many languages.