Amazon Polly can now clone a person's voice from just a few minutes of recordings, putting voice cloning within reach of businesses on Amazon's cloud, with consent checks and hidden watermarks built in.
Key Takeaways
- Voice cloning in Amazon Polly is now generally available as a gated release
- Customers can create a cloned voice from a few minutes of recordings from one consenting speaker
- The feature launches with English, Spanish, German, French, Italian, Portuguese, and Hindi
- Building a cloned voice is free, while cloned voice usage is charged on a pay-as-you-go basis
Amazon Web Services (AWS) announced that customers can give Amazon Polly a few minutes of audio from a speaker who has agreed to be cloned, then generate new speech in that speaker's voice through the Amazon Polly generative engine.
Access is already enabled for select Amazon Polly customers. Others can contact their AWS account team or apply for access. The feature is available in all AWS Regions supported by the Amazon Polly generative engine, including AWS GovCloud (US).
How Voice Cloning Works
Amazon Polly builds a custom voice from recordings of a single speaker, then lets customers use that voice for real-time audio with SynthesizeSpeech or long-form audio written to Amazon S3 with StartSpeechSynthesisTask. AWS said the cloned voice can be used the same way as other generative voices in Amazon Polly when Engine="generative" is selected.
The creation flow has four steps. Customers record one speaker, call CreateVoice with 1 to 10 recordings and the spoken language, wait while Amazon Polly builds the voice in the background, then use the returned voice ID once the status changes from pending to active or failed. Creation is asynchronous and typically takes 5 to 10 minutes.
| Voice Cloning Step | What Amazon Polly Does |
|---|---|
| Record | Capture one speaker and send recordings directly, upload them in the console, or store them in S3 in the target AWS Region |
| Create | Call CreateVoice with 1 to 10 recordings, the language, and TermsAccepted=True to confirm consent |
| Wait | Amazon Polly returns a voice ID immediately, builds the voice in the background, and typically finishes in 5 to 10 minutes |
| Use | Pass the voice ID to SynthesizeSpeech for real-time audio or StartSpeechSynthesisTask for long-form audio |
Recording quality matters. AWS recommends 2 to 10 minutes of natural speech, split into takes of up to a minute, with up to 10 recordings and up to 15 minutes of audio in total. The company recommends a quiet room, one speaker, a consistent mic distance, and mono WAV at 16-bit and 24 kHz. It also warns against meeting apps and browser recorders because echo cancellation and noise suppression can alter the voice the clone learns.
Also Read: Amazon’s Bedrock Fixes GPT Access Gap For 5 Asian Markets, No More Throttling
Consent And Controls
A cloned voice is made from one person's recordings, and customers confirm the speaker's consent each time they create a voice, whether in the console or by setting TermsAccepted=True in the API. AWS also recommends keeping a record of each speaker's consent, deleting the voice if consent is withdrawn, telling listeners when they are hearing a synthetic voice where appropriate, and limiting permissions to the people and systems that need them.
Generated speech carries a watermark. A cloned voice is scoped to the AWS account that created it, and access can be controlled with AWS Identity and Access Management. Customers can encrypt voices with their own AWS Key Management Service key, tag them, and delete them with DeleteVoice once the voice is active or if creation has failed.
AWS said Amazon Polly does not retain a copy of the voice recordings once the voice is created and does not use them to train Amazon Polly models. It keeps only the voice data needed to synthesize speech, stores that data encrypted, and says DeleteVoice removes the data immediately.
Why This Matters
Voice cloning used to sit mostly with specialist tools. Building it into one of the world's biggest clouds means a bank, a game studio or an online course maker can give its app a real person's voice in minutes, in seven languages including Hindi, without leaving the systems it already runs on.
It also raises the stakes on misuse. Cloned voices are already used in phone scams, which is why AWS ties every voice to a consent confirmation, watermarks the audio and keeps voices locked to the account that made them. Those safeguards depend on customers telling the truth about consent, so how strictly Amazon checks applications for the gated release will matter.
What To Watch Next
The feature opens up a wider range of branded and custom audio use cases inside AWS applications, from automated phone menus, voice agents, mobile apps, and in-store announcements to audiobooks, courses, podcasts, game characters, and videos. For multilingual work, there is a clear limit. Each cloned voice speaks only the language it was recorded in, so customers need to create and review approved voices for each supported language.
AWS has already enabled access for select customers, while others still need to apply or work through their account team. The company has said building a cloned voice is free for Amazon Polly customers, but generated usage is billed on a pay-as-you-go basis.
Read Next: Fish Audio’s Drama 3 Preview Tops Open Speech Chart, Fixes Pronunciation Flaw
Originally published on Fathom News.
Top comments (0)