Gemini TTS - Native Audio Out

Google’s “native audio out” for Gemini is now available in preview, letting developers generate speech directly from Gemini models with controllable voice and delivery style—single-speaker narration or multi-speaker “podcast-like” dialogue. The practical payoff is that prompts can shape not just what gets said, but how it’s performed (e.g., excited, whispering, stern), and multi-speaker configurations can produce back-and-forth conversations that resemble NotebookLM-style audio interactions.

In the walkthrough, the key capability is exposed through Google’s Gemini API: after initializing the SDK and selecting a text-to-speech-capable model, a request specifies (1) a prompt that includes both the instruction for delivery style and the actual text, and (2) a response modality set to audio. Voice selection is configured inside the speech settings, and the returned audio is delivered in the response candidates. The author notes a workflow difference: when running scripts outside of Google Colab, the audio payload sometimes needs conversion to base64, while Colab appears to handle it without extra steps. The response also includes metadata such as token counts, though pricing details for these audio tokens remain unclear during preview.

Two preview models are highlighted for native audio out: “2.5 flash” and “2.5 Pro.” The flash model is recommended for general voice quality, while the Pro model may be better when multi-speaker scenarios require more nuanced emotional direction. In AI Studio, users can audition voices before coding, and the UI supports generating both single-speaker and multi-speaker audio.

For single-speaker output, the prompt structure is central. Delivery cues like “say excitedly,” “whisper softly,” “laughing and giggling,” or “stern and more angry” can be placed directly in the prompt, and the model tends to follow those instructions—though the results can sometimes sound “overacted.” The walkthrough also flags that laughter control is imperfect and somewhat stochastic: the model may place laughter at the beginning, end, or both depending on generation randomness. Temperature can be tuned, but the behavior remains variable.

For multi-speaker audio, the workflow shifts to generating a dialogue transcript first, then converting it to speech using multi-speaker voice configuration. A sample transcript has two named speakers—David and Jenny—discussing the woolly mammoth revival, with occasional interruptions. The resulting audio alternates voices accordingly, producing a fast path from text generation to a podcast-style conversation. The author suggests that adding per-line speaking guidance can improve performance, but that doing so may increase the risk of exaggerated acting.

Finally, the discussion turns to cost and deployment tradeoffs. Pricing for native audio out in preview isn’t confirmed, raising the question of whether cloud-based generation will be economical enough versus open, locally run TTS systems. Local models remain attractive for real-time use where avoiding cloud round trips can improve speed and responsiveness.