Gemini 3.8 Flash TTS

Release Summary

Google made Gemini 3.8 Flash TTS generally available in the Gemini API on September 22, 2026. The dedicated model ID is gemini-3.8-flash-tts: it accepts text and produces audio. The companion gemini-3.8-flash-lite-tts model also reached GA that day. These are speech-generation models, separate from the general-purpose gemini-3.8-flash text-output model and from the interactive Gemini 3.8 Live models.

This release gives teams comparing AI text-to-speech tools a concrete model-selection decision. Flash TTS emphasizes expressive voice quality, character design, difficult pronunciation, and long-form dialogue. Flash-Lite TTS emphasizes high throughput, lower latency, and lower audio-output cost. Both use the same request schema, so teams can evaluate them against the same scripts and switch model IDs after migration.

What Changed

Custom voices and a larger voice library

Flash TTS supports natural-language Voice design, consent-verified Voice replication, prebuilt voices, and the extended Voices endpoint. A team can create a persistent voice persona, then reuse its voice_... ID rather than repeating a long voice description in every script. Google says Flash TTS supports 130 languages; Flash-Lite supports 101.

Line-level performance control

The 3.8 TTS schema separates the exact words to be spoken from delivery instructions. A speech_metadata block carries turn-level style and speaker information. Inline tags such as <laugh>, <sigh>, and <short pause> mark brief events within the transcript. Flash TTS also supports regional accents and IPA pronunciation overrides. This structure matters when a script has several speakers or a line must be delivered in a particular way.

More stable long-form and multi-speaker output

Google positions Flash TTS for audiobook narration, podcast production, and complex two-speaker scenes. Its model documentation describes more consistent voice identity, pacing, and room tone across extended dialogue. Google also reports improvements over Gemini 3.1 Flash TTS in long-form content and two-speaker screenplay control; those are Google-published evaluations, so production teams should test their own voices and scripts.

A production-oriented Flash-Lite option

Flash-Lite TTS is the lower-cost sibling for read-aloud features, bulk dubbing, and voice-agent cascades. Google identifies it as the recommended replacement for gemini-3.1-flash-tts-preview. It shares voice design and replication support with Flash TTS, but Google positions Flash TTS as the stronger choice for nuanced acting and broader dialect coverage.

Flash TTS vs Flash-Lite TTS

Use the same evaluation script on both models. The useful comparison is audio quality and consistency against latency and total cost for the finished clip, rather than model name alone. Both models have a free tier in Google's Standard pricing table; paid rates and usage terms should be checked before a production rollout.

Compared With Gemini 3.1 Flash TTS Preview

The earlier gemini-3.1-flash-tts-preview model already supported controllable single- and multi-speaker speech. The 3.8 release makes both TTS model IDs generally available and adds voice design and replication, an expanded voice library, and structured speech_metadata. Google specifically recommends Flash-Lite TTS as the preview model's replacement. Flash TTS is the alternative when the workload needs higher-fidelity creative output.

The request and output differences matter more than changing the model string:

Migration Guide

1. Choose gemini-3.8-flash-tts for creative quality or gemini-3.8-flash-lite-tts for scale, then test a representative set of scripts.
2. Move sustained acting direction and speaker labels into speech_metadata. Keep the transcript verbatim so the model does not speak an instruction aloud.
3. Keep brief vocal events and pauses as supported inline tags. Give every turn in a multi-speaker request an explicit speaker matching the configured voice.
4. If you use a recurring custom persona, create it with Voice design and reuse its voice ID. Voice replication requires the voice owner's verified consent.
5. Check audio handling: unary responses now default to WAV. Remove any code that manually wraps returned bytes in a WAV header, or request a supported raw format explicitly.

The official model page documents both Interactions API and GenerateContent API forms of speech_metadata. Test parser behavior, speaker labels, file headers, and long scripts before routing production traffic.

Availability, Pricing & Limits

Both models are available through the Gemini Developer API and Google AI Studio. Their GA status applies to the Developer API; as of September 24, 2026, Google Cloud Text-to-Speech's separate Gemini-TTS documentation does not list either 3.8 model ID. Google says Flash TTS is rolling out in Gemini Notebook, while Flash-Lite TTS is rolling out in Google Vids; enterprise API access in Gemini Enterprise was described as coming soon in the September 23 announcement. Voice replication in AI Studio has listed regional restrictions, so check that workflow before promising it to users in every market.

Gemini Developer API Standard paid pricing through December 31, 2026 is $0.50 per million text input tokens for either model and $9 per million audio output tokens for Flash TTS or $6 for Flash-Lite TTS. The published rates rise on January 1, 2027 to $1 per million text input tokens and $18 or $12 per million audio output tokens, respectively. Audio tokens correspond to 25 tokens per second of audio, so output duration is a major cost driver.

Flash TTS accepts text input and returns audio output, with an 8,192-token input limit and 16,384-token output limit on its model page. It does not support the Live API, function calling, grounding, or thinking. For interactive, unscripted conversations, consider the dedicated Gemini Live models instead of treating TTS as a live dialogue endpoint.

Who Should Upgrade / Who Should Wait

Choose Flash TTS if:

• Narration, character voices, regional accents, or two-speaker scenes need careful direction and stable voice identity.
• Your team can migrate to structured speech_metadata and test audio quality against actual scripts.
• The quality gain justifies the higher audio-output rate relative to Flash-Lite.

Choose Flash-Lite TTS if:

• You are replacing gemini-3.1-flash-tts-preview in a high-volume read-aloud, dubbing, or voice-agent pipeline.
• Throughput and audio-output cost matter more than the widest dialect coverage or strongest acting nuance.

Wait before switching if:

• Your client still assumes headerless PCM output or embeds spoken instructions in transcript text.
• You have not tested speaker identity, pronunciation, output duration, and billing on representative production scripts.

Sources

• Gemini API release notes
• Google's Gemini 3.8 TTS announcement
• Gemini 3.8 Flash TTS model page
• Gemini 3.8 Flash-Lite TTS model page
• Gemini TTS generation and migration guide
• Gemini API pricing