Google introduced new audio generation capabilities with the Gemini 2.5 models at the Google I/O 2025. The Mountain View-based tech giant is now letting developers and individuals test these features on its platform. The two new capabilities include native audio dialog and controllable text-to-speech (TTS) with Gemini 2.5 Flash preview. While the former can natively generate human-like audio while responding to user prompts, the latter can convert any script into conversational speech. These features are currently not available to developers via application programming interfaces (APIs).
Google Showcases Gemini 2.5 Flash’s Audio Output Capabilities
In a blog post, the tech giant detailed the features of these two audio generation modes, highlighting how developers can use them to build new experiences for people. Currently, native audio dialog can be tried out in Google AI Studio’s stream tab, whereas the TTS feature can be tested in the generate media tab within AI Studio.
Native audio dialog with Gemini 2.5 Flash preview is designed for real-time conversations between a human user and the AI. The user can either type a prompt or speak it, and the AI responds verbally. This process directly generates audio, instead of first generating text and then converting it into speech.
There are several advantages to that as well. It supports affective dialog, which means when Gemini 2.5 Flash responds to the user’s tone of voice, it can recognise the emotion behind the said words. It can understand when the user sounds scared, angry, or surprised and respond accordingly.
Apart from this, the audio generation feature can express emotions when speaking, adopt different accents and linguistic styles, can access tools such as Google Search, and supports more than 24 languages.
Coming to the controllable TTS feature, it offers multi-speaker dialogue generation, can produce emotions and accents while narrating the script, control delivery speed and emphasise pronunciation, and supports the same 24 languages and language mixing.
Google says these capabilities were assessed for potential risks across the development process. The company used both internal mechanisms as well as red teaming to find and fix any vulnerabilities. The company also highlighted that all audio outputs from these models are embedded with SynthID, its watermarking technology.
