Amazon Polly
Amazon Polly: speech APIs and reusable narration on AWS
Amazon Polly is an AWS text-to-speech service with multilingual voices, four speech engines, SSML controls and character-based pricing for apps and narration.
Paid plansAccount required

- Access
- Paid access
- Usage pricing
- USD 4 / 1,000,000 charactersStandard
- Available on
- Web, API
Speech generation for applications and reusable narration
Amazon Polly adds spoken output to applications, narration and telephone systems. Developers supply text and a compatible voice through the API, SDKs or command-line tools. The AWS console also provides a browser interface for listening to speech and saving audio files.
Standard, Neural, Generative and Long-form engines offer different voice selections and prices. Language and accent availability depend on the voice and engine, with regional availability also affecting the choice. Pronunciation lexicons and SSML help shape names, terminology and delivery; pitch controls belong to Standard voices, while Generative prosody wraps complete sentences.
Polly supports both reusable audio files and interactive speech. The bidirectional streaming API accepts text as it arrives and generates audio using Generative voices. Speech marks for synchronising text or animation are available through the Standard, Neural and Long-form engines, rather than Generative streaming.
For longer scripts, asynchronous synthesis sends the finished output to an S3 bucket. That operation has a larger text limit than SynthesizeSpeech, independently of which compatible voice engine is selected. Cached speech can be replayed without another Polly synthesis charge, while storage, delivery and other AWS infrastructure can add separate costs.
Brand Voice is a negotiated project with AWS and a voice actor, rather than a short-recording cloning tool. AWS says Polly output belongs to the customer as between the customer and AWS, with rights to supplied text still required. AWS may retain inputs for service and model improvement; organisations can use its documented AI services opt-out policy.
Best for
- Developers adding speech to AWS applications
- Publishers producing reusable audio narration
- Teams building telephone prompts and spoken interfaces
Limitations
Voice, region and SSML support vary by engine. Generative voices support bidirectional streaming but not JSON speech marks. Speech marks, storage and other AWS services can add separate costs.
Amazon Polly speech features
Four voice engines
Standard, Neural, Generative and Long-form have different voice selections, regional availability and character prices. A voice must support the engine used in the request; the same language does not imply the same options across all four.
Multilingual voices and pronunciation
Polly offers language and accent variants including US and British English, French, German, Spanish, Japanese and Mandarin. Pronunciation lexicons help with names and specialist terms; a lexicon’s language must match the selected voice.
SSML delivery controls
SSML can add pauses and adjust delivery. Standard supports pitch, rate and volume; Neural, Long-form and Generative support rate and volume without pitch. Generative prosody applies to complete sentences, and other tags have their own engine restrictions.
Bidirectional Generative streaming
StartSpeechSynthesisStream accepts text progressively over HTTP/2 and returns audio as it is generated. It currently supports the Generative engine only and does not return JSON speech marks.
Speech marks for text and animation
Standard, Neural and Long-form can return word, sentence, viseme and SSML timing metadata. Speech marks are requested separately from audio and incur their own character charges; they are unavailable with Generative voices.
Audio files and long-script tasks
SynthesizeSpeech produces MP3, Ogg Vorbis, Ogg Opus, mono PCM and mu-law or A-law audio. For longer scripts, asynchronous tasks save results to an S3 bucket and accept up to 100,000 billed characters or 200,000 total characters, including markup.
Custom Brand Voice engagements
Brand Voice is a separate project with AWS and a voice actor to create a dedicated neural voice for your organisation. Scope, pricing and timeline are negotiated; it is not a self-service short-recording cloning feature.
Amazon Polly pricing
Published USD rates per million characters are $4 for Standard, $16 for Neural, $30 for Generative and $100 for Long-form. Billing is based on characters processed, not a monthly subscription. SSML tags are excluded from billed characters. Speech-mark requests are charged separately; replaying cached audio does not add another Polly synthesis charge. Eligible new-account credits and GovCloud pricing differ.
Example plans and usage rates; additional charges may apply.
Standard
USD 4 / 1,000,000 characters
Standard engine
Plan details
Per million billed input characters; SSML tags excluded. GovCloud uses different rates.
Neural
USD 16 / 1,000,000 characters
Neural engine
Plan details
Per million billed input characters; compatible voice required. GovCloud uses different rates.
Generative
USD 30 / 1,000,000 characters
Generative engine
Plan details
Per million billed input characters; available voices and regions vary.
Technical specifications
Voice generation
| Feature | Amazon Polly |
|---|---|
| Speech languages | Multilingual voices and regional accents, including bilingual English/Hindi voices. Available voices depend on the engine and AWS region; Polly speaks the supplied language rather than translating text.Engine and region availability vary |
| Voice and pronunciation controls | SSML and pronunciation lexicons guide speech. Standard supports pitch, rate and volume; Neural, Long-form and Generative support rate and volume, not pitch. Generative prosody wraps full sentences.SSML tags differ by engine |
| Voice cloning | Brand Voice is a custom engagement with AWS and a voice actor to create an exclusive neural voice. Pricing and production timeline are negotiated; it is separate from ordinary stock-voice synthesis.Contact AWS; custom engagement |
| Audio exports and streaming | SynthesizeSpeech supports MP3/Ogg Vorbis up to 48 kHz, Ogg Opus at 48 kHz, mono PCM at 8 or 16 kHz and mu-law/A-law at 8 kHz. JSON speech marks are available on Standard, Neural and Long-form; Generative streaming does not support them.SynthesizeSpeech format-specific rates; asynchronous and bidirectional operations have their own settings |
| Commercial-use terms | Polly output belongs to the customer as between customer and AWS. Third-party input text requires appropriate rights. Replaying cached audio does not incur another synthesis charge.AWS customer agreement and input rights |
| Output sample rate (Hz) | 48000MP3/Ogg Vorbis maximum; Ogg Opus 48 kHz; not PCM or telephony |
| API access | SupportedAWS account and permissions; SDK, CLI and console access |
Amazon Polly alternatives
Other speech platforms offer managed voice APIs or browser narration studios.
Google Cloud Text-to-Speech
Cloud speech with Chirp voices, Gemini delivery prompts and model-specific billing.
Amazon Polly FAQs
How much does Amazon Polly cost?
As of 9 October 2026, published USD rates per million characters are $4 for Standard, $16 for Neural, $30 for Generative and $100 for Long-form. These are usage rates, not monthly subscription prices. Speech-mark requests add character charges, GovCloud rates differ, and other AWS services can add costs.
Is Amazon Polly free?
Eligible new AWS accounts receive $100 in Free Tier credit and can earn up to another $100 by completing activities. The free account plan ends after six months or when credits are exhausted. Credits must be used within 12 months of account creation; older account allowances differ. Polly otherwise uses paid character-based pricing.
How much text can Polly convert in one request?
SynthesizeSpeech accepts up to 3,000 billed characters or 6,000 total characters, including SSML, and its audio output is limited to ten minutes. StartSpeechSynthesisTask accepts up to 100,000 billed or 200,000 total characters and writes output to S3. The Long-form voice engine and asynchronous synthesis are separate choices.
Can Polly accept text as it arrives?
Yes. StartSpeechSynthesisStream accepts incremental text over HTTP/2 while returning generated audio. It currently supports Generative voices only. The conventional SynthesizeSpeech call also returns an audio stream, but receives its text in a single request.
Which voices support speech marks?
Speech marks are available with Standard, Neural and Long-form voices. They describe word, sentence, viseme or SSML timing for highlighting text and coordinating animation. Generative voices and the bidirectional streaming endpoint do not support JSON speech marks. Audio and speech-mark requests are billed separately.
Can I change Polly’s pitch, speed and volume?
Standard voices support SSML pitch, rate and volume controls. Neural, Long-form and Generative support rate and volume without the pitch attribute. Generative prosody must wrap complete sentences; other SSML tags can be partially supported or unavailable depending on the engine.
Can I use Polly audio commercially and replay it?
AWS states that Polly output belongs to the customer as between the customer and AWS. You need rights to third-party text supplied to the service. Generated audio can be stored and replayed without another Polly synthesis charge, subject to the AWS agreement; storage and delivery services can still have their own costs.
Does Amazon Polly clone voices?
The stock-voice service does not offer self-service cloning from a short recording. Brand Voice is a custom engagement with AWS and a voice actor to create an exclusive neural voice for an organisation. Pricing and production times are negotiated separately.
Is this your product? Claim this page to update your listing.