Amazon Polly can now clone a voice from a few minutes of audio
Amazon Polly voice cloning builds a custom voice from recordings in 5-10 minutes. Access is gated, creation is free, and the audio is watermarked.

Amazon has made voice cloning in Amazon Polly generally available as a gated release. A few minutes of consented recordings from one speaker produce a custom voice in seven languages. Creating the voice costs nothing, and only usage is billed.
How a cloned voice gets built
The process has four steps: record, create, wait, use. A developer calls CreateVoice with 1 to 10 recordings, the language spoken in them, and TermsAccepted=True, which confirms the speaker’s consent. Polly returns a voice ID starting with vc- at once and builds the voice in the background. Amazon says this typically takes 5-10 minutes, and Polly also checks the voice for speaker similarity during that time. Callers poll GetVoice until the status turns active or failed.

The finished voice runs on Polly’s generative engine and is called through SynthesizeSpeech for real-time audio, or StartSpeechSynthesisTask for long-form output written to Amazon S3, with Engine="generative". Amazon says latency matches its other generative voices.
The recording limits are specific. Amazon suggests roughly two to ten minutes of natural speech, cut into clips no longer than 60 seconds. A voice accepts at most 15 minutes in total, and console uploads are capped at 5 MB per file. Mono WAV at 24 kHz is the recommended format. Amazon warns that meeting apps and browser recorders apply echo cancellation and noise suppression, and the clone will learn those artifacts. A voice also takes one style from its recordings, so a mix of upbeat and angry takes will not yield a voice that does both.
Supported languages are English, Spanish, German, French, Italian, Portuguese and Hindi. Each cloned voice speaks only the language it was recorded in, so a brand wanting one persona in two languages needs two voices. Other accents, such as British English or Canadian French, are said to work but are less tested.
Who can use it, where and at what cost
Access is switched on per AWS account. Amazon says it is already enabled for select Polly customers, and others must contact their account team or apply. Customers can check by looking for “Voice cloning” in the console’s left panel.
It runs in every Region that supports the generative engine: US East (N. Virginia), US West (Oregon), Europe (Frankfurt, Zurich and London), Asia Pacific (Sydney, Seoul, Singapore and Tokyo), Canada (Central) and AWS GovCloud (US-West). A voice exists only in the account and Region where it was created. Serving another Region means rebuilding it from the same recordings, and Amazon notes that the speaker’s consent must cover every copy.
Each account starts with 10 cloned voices per Region, and the quota can be raised through support. CreateVoice and DeleteVoice share a limit of one request per second, so bulk jobs need a queue.
Usage is billed pay-as-you-go. The announcement states no price. According to AWS’s own pricing page, generative-engine speech is listed at $30 per million characters, with a free tier of 100,000 characters a month for the first 12 months. Whether cloned voices are billed at that exact rate is not stated in the announcement, so the Polly pricing page is the place to confirm.
Consent is a checkbox, with a watermark behind it
The safety design rests mainly on the customer’s own attestation. Creating a voice means confirming that all necessary rights and consents were obtained, and usage falls under four AWS documents: the Service Terms, the Biometric Notice and Consent Terms, the Responsible AI Policy and the Acceptable Use Policy. The announcement describes no technical step that verifies the speaker is who the customer says. Amazon advises keeping a consent record for each speaker, deleting the voice if consent is withdrawn, and telling listeners when a voice is synthetic.
Two technical safeguards are described. Polly watermarks audio made with a cloned voice, though Amazon does not say how the mark works or whether outsiders can detect it. Voices are also scoped to the creating account, governed through IAM, and can be encrypted with a customer’s own KMS key. Amazon says it does not keep the recordings once the voice exists and does not train Polly models on them. DeleteVoice removes the stored voice data immediately.
The US legal backdrop is already strict for one use the announcement highlights, phone lines. In February 2024 the FCC ruled that AI-generated voices count as “artificial” under the Telephone Consumer Protection Act, according to a summary by the law firm Cooley. Telemarketers using them need prior express written consent from the people they call. A company using a cloned voice in outbound calls therefore faces consumer-consent rules on top of the speaker’s own consent.
What changes for developers and contact centers
Polly already offered stock voices. The change is a voice that belongs to a brand or a person, produced without a studio session each time a script changes. Amazon pitches it for IVR systems, voice agents, audiobooks, courses, game characters and accessibility tools. It also advises generating fixed prompts, such as phone greetings, once, getting them approved, and serving the stored file, because identical text can come out slightly differently between requests.
Operationally, a voice is an IAM resource. A serving role can be limited to one voice, with create and delete rights kept in a separate provisioning role. Amazon also warns that an existing policy allowing polly:SynthesizeSpeech on all resources already covers every cloned voice in the account.
Cloning is a core feature of specialist services such as ElevenLabs, which we reviewed earlier. Amazon’s pitch is integration with AWS security tooling, GovCloud availability and billing alongside its other voices. The source offers no listening comparison with rivals. Claims such as “similar naturalness and accuracy” to other generative voices are Amazon’s own and were not independently tested, and no figures are given for the speaker-similarity check. For more on AWS tooling aimed at agents, see our report on AWS account creation for coding agents.
Sign Up for the Daily AI Pulse
One email a day. All the stories that matter.
By signing up, you agree to our Terms of Use and acknowledge the data practices in our Privacy Policy. You may unsubscribe at any time.
Alex CarterAI News & Big Tech CorrespondentAlex Carter writes for Aiholics, keeping readers updated on the fast-paced world of AI and Big Tech.
You may also like

Codex can now run on Amazon Bedrock, with AWS logins instead of OpenAI keys

Amazon Bedrock explained: One AWS account for many AI models, with the controls attached

Kimi K3 can now run entirely inside India on Amazon Bedrock
