Amazon Polly Adds Voice Cloning — Build a Custom Brand Voice From Minutes of Audio

Amazon’s text-to-speech service can now sound like your brand. On November 19, AWS announced that voice cloning is generally available in Amazon Polly, letting organisations create a custom synthetic voice from short recordings of a speaker — and use it across their apps, contact centres and content.
It is a big step for Polly, which until now only offered ready-made voices. Now a company can have its own voice read out its help articles, answer its support calls, or narrate its videos.
What voice cloning in Polly does
The feature runs on Polly’s Generative engine, the same engine AWS upgraded recently with more natural and expressive speech. You give Polly a few minutes of a speaker’s voice, and it builds a custom generative voice model from those recordings. That voice can then speak any text you send it, in the same tone and style as the original speaker.
AWS pitches it squarely at brands: one consistent voice across every touchpoint, instead of a different stock voice in each product. A bank’s app, its phone support and its training videos could all speak in the same voice — one that customers learn to recognise as the brand itself.
The launch supports seven languages at the start: English, Spanish, German, French, Italian, Portuguese and Hindi. That covers a large share of Polly’s audience, and Hindi in the list is a useful sign for developers building for Indian users.
How it works, and what it costs
Creating a cloned voice is simple on paper. You submit between one and ten recordings, with a total limit of 15 minutes of audio. Polly takes about five to ten minutes to build the voice model. Building the voice is free; you only pay when you generate speech with it, under Polly’s standard pay-as-you-go pricing for the Generative engine.
There are limits to know about. Each AWS account can create up to ten cloned voices per Region at launch, and a cloned voice is available only in the AWS account and Region where it was created — it does not travel with you to another region. The feature runs in the AWS Regions that support Polly’s Generative engine, including the AWS GovCloud US Region.
One detail worth noting: the recordings must come from a consenting speaker. AWS describes the capability as building voices from “short audio recordings of consenting speakers,” and that word does real work here. Voice cloning is a technology with obvious misuse potential — a voice that sounds like a real person can be used to fool people — so consent is the line AWS draws, and developers should treat it as a hard rule, not a suggestion.
The catch: it is a gated release
Here is the part that will frustrate some developers. Voice cloning is generally available, but access is gated by account. Select Polly customers already have it switched on; everyone else has to apply through their AWS account team or ask for access. It is not a button you can press in the console on day one.
That gate is probably deliberate. Cloned voices are a sensitive capability, and AWS has an interest in knowing who is building them and why. The application process also gives AWS a chance to check that the use case is legitimate before the voice exists.
If you already use Polly, the upgrade path is painless. The Generative engine upgrade that arrived alongside is a drop-in replacement: voice names and API calls are unchanged, so existing applications keep working as they do today. The upgrade also added 14 new voices across nine locales, taking the Generative engine to 57 voices across 11 languages and 25 locales, with Hindi and Finnish available in the Generative engine for the first time.
Where a cloned voice actually helps
The strongest use cases are the ones where consistency matters. A contact centre that answers thousands of calls a day can keep one voice across every call. An e-learning company can have the same instructor voice across a whole course library. A media company can narrate audiobooks or video content in a voice that belongs to it, not to a stock catalogue.
It also matters for accessibility and for agentic applications — the LLM-powered assistants that now talk to users out loud. A voice that matches the product’s character makes those conversations feel less generic, and a familiar voice can make a support call feel less like talking to a machine.
The obvious risk is impersonation, and that is why the consent requirement and the gated access matter. A cloned voice should belong to the person who recorded it, used with their permission, for the purpose they agreed to. Anything else is a misuse of the tool, and developers who cut corners there will find the consequences arrive fast.
For teams already building on AWS, Polly’s voice cloning is now the simplest path to a custom voice — no separate vendor, no new contract, just an application for access and a few minutes of clean audio. And if you are following the wider AI voice race, it sits alongside the models reshaping how machines speak: we covered Gemini 4 Neon this week, and OpenAI’s models redesigning AMD’s GPU circuits — a reminder that the same generation of models is now rebuilding both chips and voices.
