Gemini 3.8 Flash TTS

Gemini 3.8 Flash TTS (gemini-3.8-flash-tts) is Google's flagship creative text-to-speech model, engineered for studio-grade voice fidelity, expressive acting, authentic regional accents, and rock-solid long-form multi-turn stability.

Overview and capabilities

Gemini 3.8 Flash TTS sets a new benchmark for expressive audio generation:

• High acoustic fidelity and acting nuance: Delivers rich emotional range, natural cadence, and precise adherence to turn-level style directions and inline vocal events (<laugh>, <sigh>, <short pause>).
• Long-form multi-turn stability: Maintains consistent voice identity, timbre, volume, and acoustic room tone across extended dialogues and multi-minute narrations without voice drift.
• Authentic regional accents and pronunciation: Supports regional accents, minority dialects, and inline International Phonetic Alphabet (IPA) overrides.
• Full voice ecosystem support: Works seamlessly with prebuilt voices, the Extended Voice Library (GET /v1beta/voices), custom Voice design personas, and Voice replication (persistent stored voices by default, plus optional stateless keys). You can also design and replicate voices interactively in Google AI Studio.

Visit the Text-to-speech guide for full coverage of features, prompting best practices, and code examples.

When to use which TTS model

Both Gemini 3.8 TTS models share the same API schema and prompting structure. Choose the model that fits your workload:

Migration guide

If you are migrating from gemini-3.1-flash-tts-preview or earlier Gemini TTS models, update your requests for the Gemini 3.8 TTS schema:

1. Move turn-level directions into speech_metadata: Gemini 3.8 TTS treats input text strictly as a verbatim transcript. Inline text directions like "Say cheerfully: Hello!" or "Speaker 1: Hello!" may be spoken aloud. Move sustained delivery instructions (style) and speaker labels (speaker) into structured metadata:
   • Interactions API: Attach an annotation with "type": "speech_metadata", "speaker", and "style" to each text content block.
   • GenerateContent API: Attach "speech_metadata": {"speaker": "...", "style": "..."} to each part.
2. Use angle-bracket inline tags only for point-in-time vocal events: Keep momentary non-speech vocalizations and pauses inline in the transcript using angle brackets (such as <laugh>, <sigh>, <cough>, <breath>, or <short pause>). Put delivery styles like whispering in speech_metadata.style.
3. Specify speaker on every turn in multi-speaker requests: Every turn in a multi-speaker request must explicitly include speaker inside speech_metadata matching one of the configured speakers.
4. Design personas upfront with Voice design: Replace long legacy Audio Profile / Director's Notes blocks with a custom voice created in Voice design, then carry that voice_... ID through your TTS requests with minimal or empty style strings.
5. Account for default WAV (audio/wav) output on unary requests: Unlike gemini-3.1-flash-tts-preview and earlier TTS models (which returned headerless raw PCM audio/l16 by default), Gemini 3.8 TTS returns WAV audio (audio/wav / AUDIO_WAV) with a standard RIFF header by default for unary requests.
   • If your code previously wrapped raw PCM bytes in a WAV header (for example, using Python's wave module or ffmpeg), remove the manual header wrapper and write the returned bytes directly to a .wav file.
   • If your pipeline requires headerless raw PCM, mu-law, or A-law audio, explicitly set response_format to "audio/l16" ("AUDIO_L16"), "audio/mulaw" ("AUDIO_MULAW"), or "audio/alaw" ("AUDIO_ALAW"). See Audio output formats.

gemini-3.8-flash-tts

Supported languages

gemini-3.8-flash-tts detects the input language automatically and supports 130 languages: