On September 23, Google announced that Gemini 3.8 Flash TTS and Flash-Lite TTS were beginning to roll out for developers through the Gemini API and AI Studio. That does not establish immediate access for every account or region.
The documented model IDs are gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. The new request format treats input text as words to be spoken verbatim.
Delivery directions previously embedded in the transcript move to speech_metadata.style. In a multi-speaker request, each turn uses speech_metadata.speaker to identify the speaker. Leaving directions in the spoken text can cause them to be read aloud.
The response format also depends on how audio is requested. A unary request returns audio/wav by default, with a RIFF header and 24 kHz mono 16-bit PCM. Streaming instead defaults to headerless audio/l16 chunks.
Two-speaker and voice-replication limits
A single multi-speaker request supports up to two prebuilt voices. To use designed or replicated voices for several speakers, the documented path is to generate each speaker's turn separately.
Each unary WAV has a 44-byte RIFF header by default. When joining turns, request raw PCM through audio/l16 or remove each WAV header before concatenating the PCM frames.
Voice replication requires a verbal consent recording from the voice owner. Google says replication through AI Studio is not offered in Illinois, Texas, the EEA, the UK, Switzerland or India. This is not a blanket regional restriction on every TTS capability.
Audio paths to inspect during migration
A migration review can first check stage directions left in transcripts, speaker annotations and the returned audio MIME type.
Code that formerly prepended a WAV header to raw PCM must avoid adding another header to the new unary WAV. Applications needing raw samples can specify response_format.mime_type as audio/l16.
These are compatibility checks drawn from published API documentation, not audio generated or played for this article.
The first Gemini 3.8 TTS migration checks concern instruction placement and unary-versus-streaming audio formats, before any subjective sound-quality comparison.