Gemini 3.8 Flash TTS for Indian business voice calls
Google made Gemini 3.8 Flash TTS and Flash-Lite TTS generally available on 22 September 2026. For Indian businesses it adds a third credible voice option beside Sarvam and ElevenLabs, and it is most attractive when you already run Gemini as your chat model and want one provider and one bill.
- Gemini 3.8 Flash TTS and Flash-Lite TTS reached general availability on 22 September 2026.
- Flash-Lite trades quality for speed and cost, which usually suits live calls better than the larger variant.
- The strongest reason to choose it is consolidation: one Google key for both the brain and the voice.
- Sarvam still tends to lead on Indic pronunciation, so test Hindi and regional names before switching.
- General availability matters. A preview model can change or disappear, as the Gemini 2.5 access restriction showed.
What Google released
On 22 September 2026 the Gemini API changelog recorded that Gemini 3.8 Flash TTS and Gemini 3.8 Flash-Lite TTS had become generally available, described as the next generation of Google's text to speech audio models, along with a new endpoint for them.
That followed a run of releases through the month. Gemini 3.8 Flash reached general availability on 2 September, and Gemini 3.8 Live on 15 September. The voice models arriving at the end of the month complete a set: one provider covering the chat model, the live conversational model and the speech.
The phrase to notice is general availability. In the same month Google limited access to its older Gemini 2.5 models to accounts that had actively used them, which is a reminder that preview and generally available are meaningfully different commitments. Building a phone system on a preview model is building on something that can be withdrawn.
Flash and Flash-Lite, and which to use
Google ships the pair deliberately. Flash is the fuller model and Flash-Lite trades some quality for lower cost and faster response.
For live telephony, Flash-Lite is usually the better choice, for the same reason Turbo variants win elsewhere. A caller who has stopped speaking is waiting, and if the gap stretches past about a second they interrupt, assume the line dropped, or lose patience. The delay is more noticeable to a caller than the richness of the voice.
Keep the fuller model for audio generated once and played many times: a recorded brand greeting, an IVR menu, a message on hold. There the extra rendering time costs nothing because nobody is waiting on it.
A practical split is to use Flash-Lite on conversational agents and Flash on any fixed audio you pre generate. In Convo360 the voice is a per agent setting, so that arrangement needs no special configuration.
The real argument for it is consolidation
Be honest about why you would pick this over the alternatives. It is rarely because Google has the best Hindi voice.
If you already run Gemini as your chat model, using Gemini for speech as well means one provider, one API key, one bill, one set of quota limits and one changelog to watch. For a small business without an engineer, that reduction in moving parts is worth real money in avoided confusion.
The reverse argument is concentration. When a provider restricts access, raises a price or has an outage, a business running brain and voice on the same account is affected twice over. Google's own Gemini 2.5 access restriction in September 2026 is the illustration: a single decision that touched anyone depending on those models.
Neither answer is wrong. Just make the choice deliberately rather than by accident, and if you do consolidate, keep a second provider's key configured so you are one dropdown away from a fallback.
Test it the way your customers will hear it
A vendor demo reading neutral English says nothing useful about Indian calls. Build a test script from what your calls actually contain.
- Customer names: Lakshmi, Venkatesh, Shreya, Abdul.
- Local places: Kandivali, Yelahanka, Gachibowli.
- Rupee amounts spoken naturally, not digit by digit.
- A Hinglish sentence that switches language halfway, which is how most customers speak.
- A phone number or order id grouped the way a person would say it.
Listen on a mobile over a real network, not on laptop speakers. Call audio is compressed, and a voice that sounds warm on a desk setup can sound muddy on a weak signal in a shop.
Sarvam remains the specialist here. It is built for Indian languages and prices speech at around three rupees per thousand characters, so for a mostly Hindi or regional language operation it is the one to beat rather than the one to replace. Our comparison of Indian voice options sets out where each wins.
Cost is only one line of four
When comparing voice providers it is easy to focus on the speech rate alone. On an AI call that is one of four charges, and usually not the largest.
A call bills telephony minutes to your telephony provider, speech to text to your transcription provider, tokens to your language model, and synthesis to your voice provider. Halving the voice cost on a call dominated by telephony minutes barely moves the total.
So price the whole call before switching anything. Our breakdown of what an AI call really costs per minute in India works through the four components with realistic numbers.
Because Convo360 is bring your own keys, every one of those four charges goes to the provider directly at their rate, with no platform markup on minutes, characters or tokens. That makes the comparison honest, since the number in your Google or Sarvam dashboard is the number you actually pay.
Watch the changelog, not the blog
If you do adopt Gemini for voice, follow the Gemini API changelog rather than Google's AI blog. The changelog is where general availability, new endpoints and access restrictions are announced. The blog carries launches and research, and did not warn anyone about the 2.5 access change.
The same holds for whichever providers you end up with. OpenAI publishes retirements on a deprecations page its news feed does not carry. Anthropic uses API release notes. One technical page per provider, checked weekly, is the entire discipline.
Convo360 reads those pages for Meta, OpenAI, Anthropic and Google several times a day and emails the owner immediately when an entry removes a feature or changes pricing, holding everything else for a weekly digest. A voice model going away is exactly the kind of change that is silent until a caller hears nothing.
Step by step
- Decide Flash or Flash-LiteUse Flash-Lite for live conversation where the caller waits, and Flash for pre generated fixed audio.
- Write an Indian test scriptInclude customer names, local place names, a rupee amount, a phone number and one Hinglish sentence.
- Listen over a mobile networkCall your own phone rather than judging on laptop speakers, because call audio is compressed.
- Compare against SarvamFor mostly Hindi or regional calls, treat Sarvam as the benchmark to beat on pronunciation.
- Price the whole call, not the voiceAdd telephony, transcription, tokens and synthesis. Voice is often the smallest of the four.
- Keep a fallback key configuredIf brain and voice both run on Google, hold a second provider key so a switch is one dropdown.
Example conversation
Run AI calls on the voice engine you choose
Start free in minutes — no code, no credit card. Bring your own AI keys and pay no per-message markup.
Frequently asked questions
When did Gemini 3.8 Flash TTS become available?
Gemini 3.8 Flash TTS and Flash-Lite TTS reached general availability on 22 September 2026, described as Google's next generation text to speech models, alongside a new endpoint for them.
Which variant should I use for phone calls?
Flash-Lite in most cases. On a live call the pause before the agent speaks is more noticeable to the caller than the last increment of voice quality, so speed usually wins.
Is Gemini better than Sarvam for Hindi?
Sarvam is built specifically for Indian languages and generally leads on Indic pronunciation. Test both on your own names and place names before assuming either way.
Why does general availability matter?
A generally available model carries a stronger commitment than a preview one. Google restricted access to its older Gemini 2.5 models in September 2026, which showed how availability can change.
Does using Gemini for voice reduce my bill?
Possibly, but synthesis is one of four charges on a call alongside telephony, transcription and tokens. Price the complete call before switching, since telephony is often the largest line.
Related reading
Available across India — see WhatsApp & AI calling automation by state and city.