Amazon Polly has launched voice cloning as a gated, generally available feature. Customers provide two to 10 minutes of clean speech from one person, and AWS says the service usually builds a text-to-speech voice in five to 10 minutes.
The release supports English, Spanish, German, French, Italian, Portuguese and Hindi. Each clone speaks only the language in which its source audio was recorded, so a multilingual service needs a separately created and reviewed voice for each language. Access is enabled for selected customers, while others must apply or contact their AWS account team.
Creating a voice does not carry a separate fee, but synthesized speech is billed through Polly’s pay-as-you-go pricing. The voice remains scoped to the creating AWS account, can use a customer-managed encryption key and can be deleted through the API. AWS says it discards the original recordings after creation and does not use them to train Polly models.
Customers must confirm the speaker’s consent each time they create a clone, and generated audio includes a watermark. Those controls help with provenance but do not make every use appropriate. Organizations still need records of the permitted purpose, restricted IAM access and a process for withdrawal of consent. Listeners may also need clear disclosure when they are hearing synthetic speech, especially in support, advertising or personalized communications.