Gemini 3.8 TTS: voices, consent, Malayalam and API costs
Google's Gemini 3.8 Flash TTS and Flash-Lite TTS turn text into audio. The official model guide lists Malayalam for Flash TTS, while its pricing page separates free-tier use from paid API rates. Voice replication requires the voice owner's consent recording, not merely possession of a sample.
Official publisher artwork hosted by its publisher. Official source. Independent guide, not affiliated or sponsored. Artwork and trademarks belong to their owners.
Two speech models, different priorities
Google describes Flash TTS as the creative tier for voice fidelity, acting and regional dialects. Flash-Lite targets throughput, latency and cost. The current model comparison lists 130 languages for Flash and 101 for Flash-Lite.
Those are vendor descriptions, not a hands-on quality ranking. Language support does not prove that every Kerala place name, mixed Malayalam-English phrase or dialect sounds correct.
Voice design and the library
The 23 September 2026 announcement describes creating new character voices with prompts and an extended library of more than 2,000 voices. Voice remixing is labelled coming soon in that announcement.
Do not turn "coming soon" into a current button. Nor does a voice library mean you can imitate a real person without their permission.
Replication needs consent
Google describes replication from a 30-second sample of your own voice or a voice you have rights to use. It says users must supply a verbal consent recording from the voice owner that matches the reference speaker before creation.
India restriction: Google's launch page explicitly says voice replication through AI Studio is not available in India. The same restriction also names Illinois, Texas, the EEA, the UK and Switzerland. A supported language or a general Gemini API region listing does not override this feature-specific limit. Do not treat the 30-second cloning demo as a feature Indian AI Studio users can use.
The announcement describes SynthID watermarking and C2PA credentials. These safeguards do not give you rights to another person's voice or make deceptive impersonation acceptable. Tell listeners when the use could otherwise mislead them.
Current standard API prices
The live Google pricing table was checked in the browser because the fetched page omitted these rows. Standard paid rates through 31 December 2026 are US$0.50 per million input text tokens for both models, US$9 per million output audio tokens for Flash and US$6 for Flash-Lite.
At the sourced approximate Rs 96.47 per dollar reference, those are about Rs 48.24 input, Rs 868.23 Flash output and Rs 578.82 Flash-Lite output per million respective tokens. Input and audio output are different meters.
The page says the corresponding standard rates rise on 1 January 2027 to US$1 input, US$18 Flash output and US$12 Flash-Lite output, roughly Rs 96.47, Rs 1,736.46 and Rs 1,157.64 at the same illustrative rate.
These are currency estimates, not an Indian invoice. Taxes, caching, storage and other services can add costs. The page uses 25 audio tokens per second; a token price is not a flat subscription or unlimited audiobook allowance.
Free tier and private scripts
The standard table lists input and output as free of charge in the free tier, but it also marks free-tier content as used to improve products and paid-tier content as not used for that purpose. Read the applicable terms and quota before submitting private audio or scripts.
A listed free tier is not a guarantee that every account, feature, voice replication workflow or volume is eligible. We did not activate an Indian account or test actual quota.
Migration: do not speak the directions
The current model guide says input text is treated as a verbatim transcript. A line such as "Say cheerfully: Hello" may be spoken aloud. Put sustained style and speaker instructions in speech_metadata, using the current API schema.
It also says each multi-speaker turn must specify a speaker matching the configured speakers. Unary requests now return WAV by default. Older code that adds a WAV header to raw PCM should be reviewed so it does not add a second header.
Serving limits and separate products
The Flash model page lists an 8,192-token input and 16,384-token output serving limit. Google advertises long-form stability, but that is not an unlimited single request.
This model page does not support function calling or the Live API. A voice-agent system can combine services; that does not make this TTS endpoint the same as Gemini 3.8 Live.
How to try a small narration
- Open Google AI Studio through the official model page.
- Choose the exact TTS model and check quota and billing.
- Use a short script you own without personal information.
- Choose a library or designed voice first.
- In an eligible region only, use a permitted voice and complete the consent process for replication. AI Studio replication is excluded in India.
- Keep spoken text separate from delivery metadata.
- Generate a short preview and listen to every line.
- Check names, pronunciation, pauses, speaker identity and volume.
- Save the script and settings with the final audio.
What changed since the original post?
We add exact model language counts, Malayalam support, serving limits, standard API costs and the free-tier data-use distinction. The broad product announcement does not prove every account sees every feature. This revision also adds the explicit India restriction on AI Studio voice replication from the launch footnote.
No voice was cloned or audio generated for this guide. Malayalam quality and account access remain untested.
Frequently asked questions
Can I copy a voice from any recording?
No. Google requires the owner's matching consent recording.
Does Flash TTS support Malayalam?
The official language table lists it. Quality still needs listening checks.
Is this the same as Gemini Live?
No. Speech generation and real-time dialogue have separate models and pricing.
Sources: Official launch and consent safeguards · Flash TTS model and migration guide · Live standard TTS pricing table · Approximate exchange-rate reference. Checked 7 October 2026.