Amazon Polly launches voice cloning: A unique voice for your brand
Amazon Web Services launched voice cloning for Amazon Polly, enabling developers to build custom generative voices in seven languages from minutes of audio.
By Dillip Chowdary • Nov 19, 2026 • Source: AWS Machine Learning Blog

At a glance
- Product
- Amazon Polly voice cloning
- Engine
- Generative engine
- Languages
- 7 supported languages
- Creation price
- Free
- Processing time
- 5-10 minutes
- Audio limit
- 10 files, 15 minutes total
In this post
Amazon Web Services launched voice cloning in Amazon Polly as a gated general-availability release, allowing organizations to create custom generative speech models from short audio recordings of consenting speakers. According to AWS Machine Learning Blog's report, the capability operates on the Amazon Polly generative engine across all AWS Regions where that engine runs, including the AWS GovCloud US Region. Building a cloned voice is free of charge for Amazon Polly customers, while audio synthesis charges follow the standard pay-as-you-go pricing model for generative engine usage.
This breakdown covers the architectural workflow, audio ingestion requirements, supported languages, security controls, and application interfaces for developers evaluating custom voice synthesis. The release serves engineering teams building interactive voice response systems, localized narration pipelines, accessibility tools, and automated contact center workflows across AWS infrastructure.
01Seven languages and generative engine parity
Amazon Polly voice cloning launches with support for seven languages: English, Spanish, German, French, Italian, Portuguese, and Hindi. Each cloned voice speaks exclusively the language in which its source audio was recorded. The resulting custom voices run on the Amazon Polly generative engine, matching the naturalness and pronunciation accuracy of the standard stock generative voices provided by the platform.
Demonstrations from AWS staff show the system transitioning seamlessly from raw source recordings into synthetic speech reading unrecorded passages. Samples include conversational scripts, video game non-player character dialogue, interactive customer support order updates, and instructional technical narration. The engine generates dynamic speech without requiring new studio sessions when script texts change.
| Feature Attribute | Standard Generative Voices | Cloned Generative Voices |
|---|---|---|
| Voice Persona | Shared AWS catalog voices | Single consenting speaker |
| Creation Fee | $0.00 | Free |
| Engine Runtime | Generative | Generative |
| Language Binding | Multilingual catalog | Locked to recorded language |
| Initial Languages | Broad catalog support | 7 supported languages |
02Four stages of the cloning lifecycle
Creating a custom voice moves through four distinct operational phases: recording the speaker, submitting the creation job, asynchronous processing, and invoking speech synthesis.
- Record Capture audio from one speaker adhering to input specifications, sending files directly in the API request, uploading via the console, or storing them in an Amazon S3 bucket within the target AWS Region.
- Create Call CreateVoice with 1 to 10 recordings, specify the spoken language, and set the TermsAccepted parameter to True to confirm speaker consent, which immediately returns a voice identifier prefixed with vc-.
- Wait Poll GetVoice asynchronously for 5 to 10 minutes while Amazon Polly trains the voice model and runs internal quality checks for speaker similarity until the state transitions from pending to active or failed.
- Use Pass the voice identifier to SynthesizeSpeech for real-time audio playback or to StartSpeechSynthesisTask for long-form batch audio written to Amazon S3, setting Engine to generative.
Note: Access is enabled for select customers, while other AWS accounts must apply for access through their AWS account team.
03Audio specifications and capture best practices
Advertisement
Tech Pulse Daily
Get tomorrow's pulse first
Join engineers who read Tech Pulse before stand-up. Free, weekday mornings.
The synthetic model directly replicates acoustic qualities from the input audio, meaning room reverberation, microphone noise, and background computer fan hum can transfer into the generated voice. AWS specifies capturing between two and ten minutes of natural speech split into distinct takes of up to one minute each. A creation request accepts up to 10 recording files totaling no more than 15 minutes of audio.
Recordings require a single speaker positioned at a fixed distance from the microphone in a quiet acoustic environment. AWS recommends recording mono WAV files at 16-bit resolution and a 24 kHz sampling rate to prevent conversion degradation, although other common audio formats are accepted. Higher audio quality yields better similarity than simply providing larger quantities of degraded recordings.
04Access policies watermarking and encryption safeguards
Every voice creation request requires explicit consent verification by setting TermsAccepted to True in the API or confirming authorization within the AWS Management Console. Cloned voices remain strictly scoped to the AWS account that initiated creation, preventing cross-account access through standard AWS Identity and Access Management policies. Audio generated with cloned voices carries an embedded watermark to identify synthetic speech.
Amazon Polly does not retain source audio recordings once model creation completes and does not train foundational Amazon Polly models on customer voice data. The service retains only the necessary synthesis artifacts in an encrypted state, with support for customer-managed AWS Key Management Service keys. Invoking DeleteVoice removes all associated voice data immediately from the service.
05Console deployment and speech synthesis execution
Developers configure and synthesize speech using the AWS Command Line Interface or AWS software development kits by targeting the Amazon Polly API endpoints.
aws polly create-voice \
--voice-name "BrandVoiceAlpha" \
--language-code "en-US" \
--terms-accepted \
--region "us-east-1"aws polly get-voice \
--voice-id "vc-0123456789abcdef0" \
--region "us-east-1"aws polly synthesize-speech \
--engine "generative" \
--voice-id "vc-0123456789abcdef0" \
--text "Welcome to your personalized account portal." \
--output-format "mp3" \
"output.mp3"aws polly delete-voice \
--voice-id "vc-0123456789abcdef0" \
--region "us-east-1"Frequently asked questions
How much audio is required to clone a voice in Amazon Polly?
AWS recommends recording between two and ten minutes of clean speech, divided into takes of up to one minute, with a hard maximum of 10 recordings totaling 15 minutes.
Which languages are supported at launch?
The service supports English, Spanish, German, French, Italian, Portuguese, and Hindi, with each cloned voice speaking only the language used in its training audio.
How does Amazon Polly secure voice recordings and access?
Voices are scoped exclusively to the creator AWS account with IAM, encrypted using optional AWS KMS keys, watermarked during synthesis, and source recordings are discarded after creation.
Sources
Author
Dillip Chowdary
Writes Tech Bytes coverage of AI, engineering, and the tools that actually ship. Editor of Tech Pulse Daily.
Related on Tech Bytes
Advertisement