# Amazon Polly Voice Cloning Is Generally Available: Gated Access, Seven Languages, Consent Required

> Source: <https://explainx.ai/blog/amazon-polly-voice-cloning-generally-available-gated-seven-languages-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>
> Published: 2026-11-19 00:00:00+00:00

**Key takeaway:** Cloned voices are created via a CreateVoice API call with 1 to 10 recordings of up to 15 minutes total from one consenting speaker, build asynchronously in 5 to 10 minutes, are usable only by the creating AWS account via IAM, and are capped at 10 per account per Region.

Amazon Polly now offers **voice cloning** to customers. AWS announced on November 19, 2026 that the feature is generally available "as a gated release": you give Polly a few minutes of audio from a speaker who agreed to be cloned, and it builds a private custom voice that runs on the Polly generative engine. It launches in **seven languages**, building a voice is **free**, and the audio it produces carries a **watermark**. The details come from the [AWS Machine Learning Blog launch post](https://aws.amazon.com/blogs/machine-learning/amazon-polly-launches-voice-cloning-a-unique-voice-for-your-brand/?ref=explainx).

This post covers what you get, what the gate means, the exact API flow, the limits that will bite you in production, and the questions the announcement leaves open.

## TL;DR: Amazon Polly voice cloning at a glance

| Question | Answer from AWS | 
|---|---|
| Is it generally available? | Yes, as a gated release. Select customers are enabled; others apply or ask their account team | 
| What do I need? | 1 to 10 recordings, up to 15 minutes total, from one consenting speaker | 
| Languages? | English, Spanish, German, French, Italian, Portuguese, Hindi | 
| How long to build? | Typically 5 to 10 minutes, asynchronous | 
| Price? | Creating a voice is free; usage is pay-as-you-go, see the Polly pricing page | 
| Where does it run? | Every Region that supports the Polly generative engine, including AWS GovCloud (US-West) | 
| Who can use the voice? | Only the AWS account that created it, controlled with IAM | 
| Quota? | 10 cloned voices per account per Region at first | 
| Is the output marked? | Yes. Polly adds a watermark to audio from a cloned voice | 

## What is Amazon Polly voice cloning?

Voice cloning lets you build a custom Polly voice from recordings of a single speaker, then synthesize new speech in that voice through the same API you already use for the stock voices. AWS says cloned voices run on the **Polly generative engine** and deliver "similar naturalness and accuracy as the rest of the generative voices." Synthesis latency is the same as other generative voices, and the text limits follow the generative-engine quotas.

If you missed the engine itself, we covered the earlier upgrade in [Amazon Polly Generative Engine Upgrade: 14 New Voices, Hindi and Finnish](https://explainx.ai/blog/amazon-polly-generative-engine-upgrade-14-voices-hindi-finnish-same-price-2026). That post lists the generative price of $30 per 1 million characters. Cloned-voice usage is billed separately on pay-as-you-go terms, and the launch post does not repeat a number, so read the [Amazon Polly pricing page](https://aws.amazon.com/polly/pricing/?ref=explainx) before you commit.

## How does the gate work?

The word "gated" is the important one. AWS does not turn this on for every account. The launch post gives three facts:

- Access is enabled per AWS account.
- Select Amazon Polly customers already have it.
- Everyone else can reach out to their AWS account team or apply for access.

The quick test is the console. Expand the left panel in the Polly console. If **Voice cloning** is listed, it is enabled for your account. If not, it is not, and API calls to `CreateVoice` will not help you. AWS does not publish approval criteria or a timeline in the announcement, so treat the wait as unknown.

## How to build a cloned voice with the API

AWS describes a four-step lifecycle: record, create, wait, use. Its own figure shows the flow.

Amazon Polly voice cloning lifecycle from recording a speaker to creating, waiting for and using the cloned voice
*Image: [AWS](https://aws.amazon.com/blogs/machine-learning/amazon-polly-launches-voice-cloning-a-unique-voice-for-your-brand/?ref=explainx)*

1. **Record.** One speaker, one quiet room. You can send recordings inside the`CreateVoice` request, upload them in the console, or point Polly at an Amazon S3 bucket in the Region where the voice will live.
2. **Create.** Call`CreateVoice` with 1 to 10 recordings, the language code and`TermsAccepted=True` . Polly returns a voice ID that starts with`vc-` and builds in the background.
3. **Wait.** Creation takes 5 to 10 minutes. Polly also checks the quality of the voice for speaker similarity during this time. Call`GetVoice` until the status moves from pending to active or failed.
4. **Use.** Pass the voice ID to`SynthesizeSpeech` for real-time audio, or to`StartSpeechSynthesisTask` for long audio written to S3. Set`Engine="generative"` . Call`DeleteVoice` when you are done.

The AWS example in Python uses boto3. This is the shape of the first call, adapted from AWS's sample:

``` python
import boto3

polly = boto3.client("polly", region_name="us-east-1")

voice = polly.create_voice(
    Name="MyBrandVoice",
    LanguageCode="en-US",
    TermsAccepted=True,  # confirms the speaker's consent and the terms of use
    VoiceCreationData={
        "Recordings": [
            {"Format": "wav", "Data": {"S3Url": "s3://amzn-s3-demo-bucket/brand-voice/take-1.wav"}},
        ],
    },
    Tags={"consent-record": "speaker-0042"},
)["Voice"]
print(voice["Id"], voice["Status"])
```

You need the latest boto3. If Python reports that the Polly client has no attribute `create_voice`, upgrade the library. Permissions needed are `polly:CreateVoice`, `polly:GetVoice`, `polly:DescribeVoices`, `polly:SynthesizeSpeech` and `polly:DeleteVoice`, plus `polly:StartSpeechSynthesisTask`, `polly:GetSpeechSynthesisTask` and `s3:PutObject` for long-form audio.

## How should I record the source audio?

AWS spends real space on recording advice, because the clone learns everything in the file. Its guidance, in order of how much it will matter:

- **Two to ten minutes of natural speech** , split into takes of up to a minute. Up to 10 recordings and 15 minutes in total. AWS says quality matters more than length.
- **One speaker, a quiet room, a constant distance from the mic.** Use a USB microphone or a phone app that records WAV. Mono, 16-bit WAV at 24 kHz is the recommended format. Polly accepts other formats too.
- **Do not use meeting apps or browser recorders.** Their echo cancellation and noise suppression change the voice, and the clone learns those changes.
- **Record in the style you want.** The voice takes its style from the recordings and cannot blend several. Upbeat, sad and angry takes will not produce a voice that does all three.
- **Record in the language the voice should speak.** AWS tested en-US, de-DE, es-US, fr-FR, it-IT, pt-BR and hi-IN. Other accents such as British English or Canadian French work but are less tested, so check them against your own content.

Console uploads are limited to 5 MB per file. For larger files use S3.

## Consent, watermarking and data handling

This is the part of the launch with the most weight, because voice cloning is also the tool behind scams and impersonation. For context on the risk side, see our posts on a [$25 million deepfake video-call scam](https://explainx.ai/blog/deepfake-fraud-25-million-video-call-scam-verification-2026) and on [synthetic voices in 911 swatting calls](https://explainx.ai/blog/ai-synthetic-voices-911-swatting-apco-survey-november-2026).

What AWS puts in place:

- **Consent per voice.** You confirm, on the console or with`TermsAccepted=True` , that you hold all rights and consents to create and use the voice. Use is subject to the AWS Service Terms, the Biometric Notice and Consent Terms, the AWS Responsible AI Policy and the AWS Acceptable Use Policy.
- **A watermark.** Polly adds a watermark to text-to-speech audio generated with a cloned voice. The launch post does not describe how it works or how to detect it, so do not assume a public detector exists.
- **Account scoping.** A cloned voice belongs to the account that created it. You control access with IAM, and you can encrypt with your own AWS KMS key.
- **No recording retention.** AWS says Polly does not keep a copy of your recordings once the voice exists and does not use them to train Polly models. It keeps only the encrypted voice data needed to synthesize speech, and`DeleteVoice` removes that data immediately.

The consent step is a checkbox-style attestation. Polly does not verify who the speaker is from the API alone. That is why the AWS recommendations matter: keep a record of each speaker's consent (the sample code stores a consent record ID as a tag), delete the voice if the speaker withdraws consent, tell listeners when they hear a synthetic voice where appropriate, and limit permissions to the people and systems that need them.

## What can you build with it?

A speech bubble sending sound arcs to a speaker, standing for text turned into a branded cloned voice in Amazon Polly

AWS lists five target uses:

| Use | What AWS says | 
|---|---|
| Brand voice across channels | One voice for IVR, voice agents, mobile apps, in-store announcements, ads and connected devices | 
| Contact centers | Update prompts when offers or policies change without a new studio session | 
| Media at scale | Audiobooks, courses, podcasts, game characters and videos in multiple custom voices | 
| Localization | Keep a recognizable persona across a library, with one approved voice per language | 
| Accessibility | Natural-sounding voices for reading tools and personalized interactions | 

The launch post includes sample pairs: a real recording followed by the cloned voice, including a German speaker and a game non-player character. They are demos, not benchmarks. AWS publishes no speaker-similarity score or listening-test result in the announcement, so the quality claim is unverified beyond those clips.

For comparison with the rest of the market, explainx.ai has covered open and commercial cloning options: [Edge0 Audio8, an open 0.6B model with voice cloning in 23 languages](https://explainx.ai/blog/edge0-audio8-tts-realistic-air-0-6b-open-voice-cloning-23-languages-apache-2-0-2026), [VoxCPM2](https://explainx.ai/blog/voxcpm2-tokenizer-free-tts-multilingual-voice-cloning-2026), [Voicebox](https://explainx.ai/blog/voicebox-open-source-ai-voice-studio-elevenlabs-alternative-2026) and [ElevenLabs Eleven v4](https://explainx.ai/blog/elevenlabs-eleven-v4-v4-turbo-voice-models-2026). Microsoft also gates instant voice prompting from a short consented clip, as we noted in the [MAI-Voice-2-Flash launch](https://explainx.ai/blog/microsoft-mai-image-2-5-pro-voice-2-flash-foundry-2026). The pattern is clear: the big clouds ship cloning behind access gates and consent terms, while open models ship it with a license and nothing else.

## Production limits to plan for

The "Things to know" section of the launch post holds the details that decide whether this fits your system.

- **Quota.** 10 cloned voices per account per Region at first. You request more through your account team or AWS Support.
- **Rate limit.**`CreateVoice` and`DeleteVoice` share a limit of 1 request per second, so queue bulk jobs.
- **Regional scope.** A voice exists only in the account and Region where you created it. To serve another Region, create it again from the same recordings. AWS warns you to make sure the speaker's consent covers every copy.
- **Provisioning, not request time.** Creation takes minutes, so treat it as a setup step, for example in a Step Functions workflow, and store the voice ID in config.
- **Retries can duplicate.** If`CreateVoice` times out, the voice may already exist and an SDK retry can create a second one. Check`DescribeVoices` for the name first. Cloned voices show`Origin` set to`customer` .
- **Wildcard IAM policies.** A statement that allows`polly:SynthesizeSpeech` on`"Resource": "*"` also covers every cloned voice in the account. Review old policies and scope serving roles to a single voice ARN.
- **Output varies.** The same text can come out slightly differently between requests. For fixed prompts such as phone greetings, generate once, get approval from the speaker and your brand team, then serve the stored file.
- **Pronunciation.** Fix mispronounced names by respelling or with an SSML phoneme tag. Phoneme tags work with cloned voices in English, French, Italian, German and Spanish.
- **Failure messages are vague.** When creation fails,`StatusReason` gives only a general message. Check recordings against the advice above, delete the failed voice and try again.

## What this means for what you build or pay

For builders, the change is small in code and large in process. The API is four calls plus the synthesis calls you already use. The work is in consent records, IAM scoping and recording quality. If you already run Polly for IVR or narration, voice cloning replaces the studio booking each time a script changes. If you do not have access yet, apply now, because the gate is the slowest step.

For founders and operators, the free build and per-use billing mean the cost shows up in usage, not setup. Because AWS has not put a cloned-voice rate in the launch post, get the number from the pricing page or your account team before you promise a margin. Also budget for the legal side: biometric consent terms apply, and a departing speaker can force you to delete the voice.

For marketers, the draw is on-demand audio in a consistent brand voice. The risk is trust. AWS itself recommends telling listeners when they hear a synthetic voice where that is appropriate. Disclosure expectations differ by market, so ask legal before a cloned voice fronts an ad.

## What is still unknown

- **Approval criteria.** AWS does not say who gets access or how long it takes.
- **Cloned-voice price.** Only "pay-as-you-go" is stated in the launch post.
- **Quality evidence.** There are no similarity scores, listening tests or comparisons to rivals.
- **Watermark detail.** Nothing on its robustness or how a third party would detect it.
- **More languages.** Finnish, Korean and others that Polly supports on the generative engine are not in the cloning list yet.

The official answers will appear in the [Amazon Polly developer guide](https://docs.aws.amazon.com/polly/latest/dg/what-is.html?ref=explainx), the pricing page and the Polly AI Service Card that AWS links from the launch post.

## Voice cloning versus the other ways to get a custom voice

Before you apply, decide whether cloning is the right tool. Three routes exist, and they trade control against effort.

**Stock generative voices.** Polly already has 57 generative voices across 11 languages. If a brand does not need a unique sound, a stock voice avoids consent paperwork and the access gate. This is the route to try first for prototypes.

**Cloned voice from a consenting employee or spokesperson.** This is what the new feature serves. The cost is one recording session and a consent record. The risk is dependence on one person: if they leave and withdraw consent, you delete the voice and rebuild the brand audio.

**Open models you host.** Models with cloning built in give you full control and no gate, but you carry the safety work yourself: consent, misuse prevention and any watermarking. Our guides on [Voicebox](https://explainx.ai/blog/voicebox-open-source-ai-voice-studio-elevenlabs-alternative-2026) and [Edge0 Audio8](https://explainx.ai/blog/edge0-audio8-tts-realistic-air-0-6b-open-voice-cloning-23-languages-apache-2-0-2026) cover that path.

A practical rule: use Polly cloning when you already run on AWS, need IAM and KMS controls, and want the vendor to carry part of the compliance burden through its terms and watermark. Use an open model when you need languages Polly does not cover yet or need to run offline.

## A short checklist before you apply

1. Pick the speaker and get written consent that names the use, the channels and the term. Keep it with the voice ID.
2. Record in one session, mono WAV, quiet room, the style you want the voice to keep.
3. Decide where the voice will run. Plan for one voice per Region and per account.
4. Write an IAM policy that limits `SynthesizeSpeech` to that voice and keeps`CreateVoice` and`DeleteVoice` on a separate provisioning role.
5. Decide a disclosure policy for listeners and write down the deletion process if the speaker withdraws consent.
6. Test with your hardest real scripts: product names, numbers, addresses, long sentences.

If you want to build voice agents on top of this, our [Claude voice data guide](https://explainx.ai/blog/claude-voice-data-training-opt-in-anthropic-settings-guide-2026) and the [Baroness Campbell case](https://explainx.ai/blog/baroness-campbell-ai-voice-house-of-lords-elevenlabs-first-speech-2026) show two sides of consented voice use: privacy settings on one side, an assistive-voice first on the other.

## Related reading

*Specifications, limits and availability are as stated by AWS on November 19, 2026 and may change. Check the Amazon Polly documentation for current details.*
