AWS Now Lets Polly Clone a Voice: a Few Minutes of Audio Buys You a Brand Voice, but You Still Have to Ask for Access
On November 19, AWS made voice cloning in Amazon Polly generally available — with an asterisk. The launch is a gated release: some Polly customers already have it switched on for their accounts, everyone else has to go through their AWS account team or an application form. Once you are in, you hand Polly one to ten recordings from a single consenting speaker, up to 15 minutes of audio in total, and it builds a private voice in the background in about five to ten minutes. After that you pay only for the speech you synthesize.
Creating a cloned voice is free. Synthesis is billed at Polly's generative rate of $30 per million characters, and new accounts get 100,000 generative characters a month free for their first 12 months.
TL;DR
- What it is: Polly is AWS's managed text-to-speech service. It can now build a private custom voice from a speaker's recordings and run it on the same generative engine as its stock voices.
- Why it matters: Rewriting a script used to mean booking a studio again. Now it means re-synthesizing. For brand voices, contact-center prompts, audiobooks and courses — cases where the words keep changing but the voice must not — the cost structure changes outright.
- What it costs you: A cloned voice speaks only the language it was recorded in, carries only the delivery style it was recorded in, and lives only in the account and Region that created it. It is not a universal voice actor.
How it works
AWS lays it out in four steps, and it looks like any other API call:
- Record. One speaker, one quiet room, the same distance from the mic every take. You can upload through the console, attach the audio directly in the
CreateVoicerequest, or point Polly at an S3 bucket in the target Region. Uploads cap at 5 MB each. - Create. Call
CreateVoicewith one to ten recordings, the language they are spoken in, andTermsAccepted=True— your declaration that you hold the speaker's consent. The call returns avc-voice ID immediately and the build runs asynchronously. - Wait. Typically five to ten minutes. During that window Polly also checks speaker similarity. Poll
GetVoiceuntil the status flips frompendingtoactiveorfailed; AWS suggests a 20-minute timeout. - Use. Pass the voice ID to
SynthesizeSpeechfor real-time audio, or toStartSpeechSynthesisTaskto render long text into S3, withEngine="generative". CallDeleteVoicewhen you are done.
One code-level gotcha worth pinning up front: if boto3 throws Polly object has no attribute create_voice, that is not a permissions problem, it is an old SDK. Run pip install --upgrade boto3.
The recording step is where most clones get ruined
There is a line in the AWS blog worth copying verbatim: a clone learns everything in the recording, including the parts you would rather it did not — the laptop fan, the room echo, the meeting app's noise filter.
Hence the specifics. Record two to ten minutes of natural speech, split into takes of no more than a minute. Use mono WAV, 16-bit, 24 kHz, and record it that way from the start so nothing needs converting later. Skip meeting apps and browser recorders entirely: their echo cancellation and noise suppression reshape the voice, and the clone learns those changes as part of the person.
The constraint people miss most often: a clone's delivery style comes from the recordings and cannot be blended. Record upbeat, sad and angry takes separately and you do not get one voice that can do all three. You get three voices that can each do one.
Where the gate actually bites
Seven languages at launch: English, Spanish, German, French, Italian, Portuguese and Hindi. Each cloned voice speaks only the language it was recorded in, so multilingual work means building and reviewing one voice per language.
Regions follow Polly's generative engine: US East (N. Virginia), US West (Oregon), Frankfurt, Zurich, London, Sydney, Seoul, Singapore, Tokyo, Canada (Central), plus AWS GovCloud (US-West).
Access is switched on per AWS account. Select Polly customers already have it; the rest contact their account team or apply. According to Unite.AI's read of the release, new accounts start with a quota of 10 cloned voices per account per Region, and CreateVoice plus DeleteVoice share a ceiling of one request per second — bulk creation and bulk deletion both need a queue.
Unite.AI also flags a hard constraint: a voice is locked to the account and Region that created it. Serving another Region or another account means rebuilding from the same recordings, with the speaker's consent re-confirmed for every copy.
To be clear about sourcing: neither the quota nor the rate limit appears on the AWS announcement page or the blog post. Those two operational numbers are compiled from multiple media reports; treat what your own account shows as authoritative.
What it costs
- Building a voice: free.
- Synthesis: the generative rate, $30 per million characters, metered per character with no upfront commitment.
- Free tier: 100,000 generative characters per month for the first 12 months.
- In practice: a typical 6,500-character news article, roughly nine minutes of audio, costs about $0.20. A 165,000-character audiobook like *A Christmas Carol* — about three hours fifty minutes — runs about $4.95. A 600,000-character title like *Huckleberry Finn*, close to fourteen hours, lands near $18.
Who it is for
- Brand consistency: one recognizable voice across IVR, apps, in-store announcements, advertising and connected devices, with no studio session when the copy changes.
- Contact centers and voice agents: self-service prompts change with every campaign and policy update; regenerate them in the same voice instead of rebooking talent.
- Long-form production: audiobooks, courses, podcasts and game characters, where later chapters can reuse the same voice.
- Accessibility and personalization: a more natural voice for reading tools and assistive software.
When it is not a good fit
- You want one voice that does many languages and many moods. One clone equals one language and one style. It cannot.
- You need the voice available across Regions or accounts on demand. It does not travel; every copy has to be rebuilt.
- The source audio is rough. Background noise gets learned. The clone is not a repair tool.
- You were hoping it replaces talent consent. AWS requires a consent confirmation on every creation and recommends keeping records and deleting the voice if the speaker withdraws consent.
Pitfalls
- Do not record through a meeting app. Noise suppression and echo cancellation alter the timbre, and the clone adopts the alteration.
- Do not assume more audio is better. AWS says quality beats length; the ceiling is 10 takes and 15 minutes total, and going past it buys nothing.
- Do not put voice creation in a live request path. Five to ten minutes per build makes it a provisioning step, not a user action.
- Do not forget the watermark and the permissions. Output from a cloned voice is watermarked, the voice is scoped to the creating account, can be governed with IAM, can be encrypted with your own KMS key, and
DeleteVoicewipes the voice data immediately. - Do not lose the consent trail. Cross-Region copies and shared access both require re-confirming the consent scope — the point where compliance usually stops these projects.
Next steps
- First, check that your Region is on the supported list, then check whether your account is already enabled. If not, go through your account team or the application form.
- Second, record a test set to spec: two to ten minutes, mono WAV, 16-bit, 24 kHz, quiet room, consistent mic distance.
- Third, run
CreateVoice → GetVoice → SynthesizeSpeechon roughly 6,500 characters of your real copy and measure both the actual unit cost and how it sounds to your ears. - Fourth, put voice creation on your pre-launch checklist, add
DeleteVoiceand consent-record retention to the runbook, and only then consider production.
Sources
- AWS What's New: Amazon Polly now supports voice cloning — https://aws.amazon.com/about-aws/whats-new/2026/11/amazon-polly-now-supports-voice-cloning/
- AWS Machine Learning Blog: Amazon Polly launches voice cloning: A unique voice for your brand — https://aws.amazon.com/blogs/machine-learning/amazon-polly-launches-voice-cloning-a-unique-voice-for-your-brand/
- Unite.AI: Amazon Polly Adds Gated Voice Cloning for Custom Brand Voices — https://www.unite.ai/amazon-polly-adds-gated-voice-cloning-for-custom-brand-voices/
- Amazon Polly pricing — https://aws.amazon.com/polly/pricing/