Gemini 3.8 TTS separates voice identity from delivery controls, requiring audio format updates to existing pipelines

Google announced Gemini 3.8 Flash TTS and Flash-Lite TTS on September 23. The former focuses on voice design and performance controls, while the latter targets high-volume speech generation. Developers should note the clearer separation of interfaces for character voices, per-turn delivery, and text to be read aloud, along with migration requirements for existing audio pipelines. [Official announcement](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-text-to-speech/)

The new version treats input text as a script to be read aloud. Emotion, speaking rate, and intonation that apply throughout a turn now belong in `speech_metadata.style`, while momentary events such as coughs, sighs, and pauses are specified using inline tags. If older prompts mix instructions such as “say this in a happy voice” into the body text, those instructions may themselves be read aloud. The default format for standard non-streaming requests also changes from headerless PCM to WAV, so existing code must remove any step that manually adds a header. Pipelines requiring raw audio should explicitly specify the format. [Migration documentation](https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash-lite-tts)

Voice design has a separate resource workflow: first call `voices.create` to create a character voice, receive a voice ID and preview audio, and then reuse that voice in subsequent synthesis requests. This lets developers manage a consistent voice identity separately from per-utterance performance parameters, making character settings easier to track. However, stored voices are limited to 200 per project and have a one-year lifetime, so lifecycle management remains necessary. [Voice design documentation](https://ai.google.dev/gemini-api/docs/voice-design)

Multi-speaker dialogue has clear limits: a single request supports at most two speakers, using preset voices only. Dialogue involving custom or cloned voices requires synthesizing each turn separately and then concatenating the 24 kHz PCM audio. In deployment, applications still need to manage transitions between turns, request scheduling, and playback buffering. [Speech generation limitations](https://ai.google.dev/gemini-api/docs/speech-generation#limitations)

Voice cloning requires both a reference recording and a recording of a specified consent statement from the same adult speaker. The documentation specifies a reference clip length of 10 to 30 seconds. This directly affects recording collection and cataloging workflows, and existing voice libraries must also be checked for recordings that meet the interface requirements. [Voice replication interface documentation](https://ai.google.dev/gemini-api/docs/voice-replication)

Flash-Lite’s language table lists Traditional Chinese and Cantonese, but language coverage does not imply uniform quality across all accents, proper nouns, and long-form content. [Model specifications](https://ai.google.dev/gemini-api/docs/models/gemini-3.8-flash-lite-tts) The next step for Chinese-language services should be to test actual scripts for mixed Chinese and English, number pronunciation, and voice consistency, while measuring time to the first audio chunk and pauses after concatenation. Testing should also cover differences before and after prompt migration, to avoid mistaking format compatibility for identical reading behavior. No independent listening tests were conducted for this article.