Sarvam's new Dubbing and Voice Cloning APIs explained

Updated 2026-10-05 · 5 min read · By the Convo360 team

What Sarvam actually shipped

On 26 September 2026 Sarvam's changelog recorded two releases on the same day, both under its Content Studio product.

The Dubbing API localises audio and video programmatically while preserving each speaker's voice. The distinction from ordinary dubbing is the voice preservation: the output is your speaker talking in another language, not a different narrator reading a translation.

The Voice Cloning API generates speech in a cloned voice from a short reference recording. That capability existed in products before this, but exposing it as an API is what makes it usable inside a workflow rather than only in an interface.

Both sit alongside Sarvam's existing speech to text and text to speech services, which is the context that matters for Indian businesses: one provider, built for Indic languages, now covering transcription, synthesis, cloning and dubbing.

Why voice preservation changes the economics of dubbing

Consider a coaching business with forty hours of recorded lessons in English that it wants to offer in Hindi and Tamil.

The traditional route is to hire voice artists, book studio time, and re record. The cost scales with hours and languages, and the result is a different person teaching, which breaks the connection students had with the original instructor.

With voice preserving dubbing, the same instructor appears to teach in Hindi. The brand relationship survives the translation, which for education, training and founder led content is most of the value.

The honest caveat is that this is not a finished product out of the box. Machine dubbing still mistimes some sentences, mispronounces names, and occasionally flattens emphasis in a way a listener notices. Treat the output as a strong first pass that a human reviews, not as a final cut. That is still a large saving on forty hours of studio time.

Where cloning earns its place

Cloning matters when the same voice says many different things, which is exactly the shape of an AI phone agent.

A dealership calling leads, a clinic confirming appointments, an NBFC chasing a payment: each call is different text in the same voice. If that voice belongs to the business, through its founder or a familiar agent, the calls sound like the company rather than like software.

Convo360 already supports cloning a voice from one or two minutes of audio and using it on calls, so a better underlying cloning model raises the quality of something already in place rather than enabling a new feature.

What cloning does not fix is script quality. A cloned founder voice reading a badly written opening line sounds like the founder saying something awkward, which is worse than a neutral voice doing it. Get the first two sentences right before worrying about whose voice delivers them.

What it costs, and how to think about it

Sarvam publishes its rates, which makes planning straightforward. Text to speech is priced at around three rupees per thousand characters, and speech to text at around thirty rupees per hour, with diarisation charged higher. Its own language models are priced per million tokens, well below frontier model rates.

For voice work the character based pricing is the one to model. A thousand characters is roughly one hundred and fifty spoken words, or about a minute of speech. So a two minute call costs a few rupees of synthesis, and the telephony minute is usually the larger line on the bill.

Because Convo360 is bring your own keys, those charges go to Sarvam directly at Sarvam's rates. There is no platform markup on characters, minutes or tokens, which means the cost of experimenting with dubbing or cloning is the provider's metered price and nothing else. Our breakdown of Sarvam pricing for Indian businesses works through the arithmetic.

Consent is not a formality

A cloned voice is a person's likeness. Before cloning anyone, get their explicit permission in writing, and be specific about what it will be used for.

The questions worth settling up front are simple and are much harder to answer after the fact. Who owns the clone if that person leaves the company. Can it be used for outbound sales or only for service messages. What happens to it if they withdraw consent. Will callers be told they are speaking to an AI.

That last point is not only courtesy. A cloned human voice is convincing enough that a caller may reasonably believe a person is on the line, and in India a call that sounds like a real person while selling something sits close to rules you do not want to test. Say plainly that it is an automated assistant, early in the call.

Keep DND and consent handling in place as well. A better voice does not change who you are allowed to call.

A sensible way to start

Pick one small, low risk piece of work rather than a programme.

For dubbing, take a single five minute video, dub it into one language, and have a native speaker review it line by line. You will learn more from that review than from any vendor comparison, because the failures are specific to your content: your product names, your numbers, your speaker's pace.

For cloning, record two minutes of clean audio in a quiet room, clone it, and use it on a non critical agent such as an internal test line or an appointment reminder. Listen to a full call on a mobile network before putting it in front of customers.

In both cases keep the original. A dubbed file and a cloned voice are derived assets, and you will want to regenerate them when the models improve, which on current evidence will be within a few months.

Step by step

1. Pick one short piece of contentStart with a single five minute video rather than a whole library, so review stays manageable.
2. Have a native speaker review itCheck line by line for timing, product names and numbers. Machine dubbing is a first pass, not a final cut.
3. Record two clean minutes for cloningUse a quiet room and a consistent microphone. Reference quality sets the ceiling on the clone.
4. Get written consent firstAgree what the cloned voice may be used for, who owns it, and what happens if consent is withdrawn.
5. Test on a non critical agentTry the cloned voice on an internal line or reminder agent before putting it on sales calls.
6. Keep the source filesRetain original audio so you can regenerate dubs and clones as the underlying models improve.

Example conversation

Frequently asked questions

Related reading

Available across India — see WhatsApp & AI calling automation by state and city.