Google Launches Gemini 3.8 Flash TTS With Voice Design and Replication

Two Gemini TTS model IDs are available now

On September 23, 2026, Google introduced Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS. The developer model IDs are gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts. Google’s launch page says both started rolling out that day in the Gemini API and Google AI Studio. Flash also appears in Gemini Notebook, while Flash-Lite appears in Google Vids. API access through Gemini Enterprise was described as coming soon rather than available now.

This is a current API launch, not only a research demo. The speech documentation provides request examples through the Interactions API and a supported-model matrix for both IDs. AccessAllGPT did not authenticate an account, however, so availability in a specific project, country, quota tier or enterprise contract remains to be confirmed in that account.

Flash and Flash-Lite share the API but target different jobs

Both models use the same schema, making the model ID the principal switch. Google positions Flash for maximum acoustic fidelity, acting control, difficult pronunciation, regional or minority dialects and long-form work. It positions Flash-Lite as the faster, lower-cost option for bulk production, read-aloud features and conversational voice-agent cascades. These workload claims come from Google and were not independently reproduced here.

The model card says both accept text up to 8K tokens and produce audio with an output ceiling of 64K tokens. The API documentation describes more than 130 supported languages for Flash and more than 100 for Flash-Lite. A language appearing in the table is not evidence that one voice, dialect, pronunciation set or safety behavior meets a production bar; test every required locale with the intended script distribution.

Launch pricing is temporary and doubles on January 1

Google’s pricing page lists the same text-input rate for both models through December 31, 2026: $0.50 per million tokens. Audio output costs $9 per million tokens for Flash and $6 for Flash-Lite. Google defines audio as 25 tokens per second and translates those rates to $0.00225 and $0.0015 per ten seconds respectively. The free tier is listed as free of charge, with free-tier data marked as used to improve Google products and paid-tier data marked as not used for that purpose.

The table already publishes the step-up: on January 1, 2027, text input becomes $1 per million tokens, Flash audio becomes $18 and Flash-Lite audio becomes $12. Input-cache and storage prices also double. A pilot approved on the launch tariff therefore needs two forecasts—one for the remaining 2026 window and one for steady-state 2027—plus retries, stored custom voices, preprocessing, delivery and application costs.

The output contract changes between unary and streaming requests

Unary requests return a complete 24 kHz mono WAV file with a RIFF header by default. Streaming requests instead return headerless 16-bit little-endian PCM chunks at 24 kHz. The API can also return linear PCM, μ-law or A-law and accepts an explicit sample rate. This matters during migration: code written for an older preview model may add a WAV header itself and corrupt an already wrapped Gemini 3.8 response.

For two-speaker audio, the documentation expects separate text items with speech_metadata identifying each speaker, and conversational mode for turn-taking. Sustained direction belongs in speech_metadata.style; momentary breaths, laughs or pauses belong as inline tags. Keep the transcript field verbatim. Add an automated format probe and listenable fixture before switching an existing production pipeline.

Voice design and replication are API resources, not just prompts

Gemini 3.8 TTS can use 30 named studio voices, an extended catalog, a designed voice created from a natural-language description, or a replicated voice created from recordings. Google’s documentation says stateful designed and replicated voices use voice_ IDs, share a 200-voice project quota and have a one-year retention period. Optional stateless replicated voicekey_ values are client-managed and expire after seven days.

Google says replication starts from a 30-second reference and requires a verbal consent recording from the same speaker. The launch post also says generated audio receives SynthID and replicated voices carry C2PA credentials. Those are vendor-described safeguards, not proof that every downstream player preserves provenance or that the requester owns the voice. Applications still need identity checks, scoped authority, visible synthetic-audio disclosure, restricted creation access, revocation and deletion.

Google reports strong voice rankings, but buyers need their own test

Google says Flash scored 71.4 on Hume AI’s Voice Design Benchmark and 60.8 for accent modeling, and says Flash and Flash-Lite ranked first and second on Hume’s Overall Quality Index. It also reports leading positions in selected languages on Voice Arena. These are vendor-selected and vendor-reported results. AccessAllGPT did not inspect evaluation samples, reproduce scores or establish that the tests represent a buyer’s voices and scripts.

A useful comparison fixes the transcript, voice authority, style instruction, output format and volume across Flash, Flash-Lite and the incumbent. Blind raters should score intelligibility, pronunciation, identity consistency, expressiveness and artifacts. Measure first-audio latency, completion latency, retries, billed tokens and accepted-audio minutes. Include difficult names, numbers, code switching, long passages, pauses and malformed markup—not just polished demos.

The practical decision is Flash, Flash-Lite or wait

Pilot Flash when a creative workflow needs designed characters, nuanced acting, long-form stability or the broader documented language set. Pilot Flash-Lite first when the main constraint is throughput or cost and the workload uses common languages with repeatable scripts. Because the schema is shared, route a frozen sample to both and choose on accepted output rather than the model name.

Wait when the required Gemini Enterprise API route, country, quota, voice-consent process or post-2026 budget is unresolved. Reject replication for any workflow that cannot prove speaker authority or revoke and delete a custom voice. Keep a non-cloned catalog voice or a text fallback ready. The API is available now; voice rights, workload quality and 2027 economics are still deployment decisions.

Disclosures

AccessAllGPT did not receive Google credits, an API account, early access, custom voices, a product briefing, test data, review or compensation for this article. Google and Google DeepMind did not sponsor, review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Google, Google DeepMind, OpenAI or organizations cited. Publication-wide relationships are listed on the disclosures page.