# Amazon Polly Generative Voices: Hindi, Finnish and Unchanged Pricing

AWS upgraded Polly's generative engine and added 14 voices, but voice quality, regional access and SSML support still need separate checks.

By Paras Madan | 2026-10-23


AWS upgraded Amazon Polly's generative text-to-speech engine on October 22 and added 14 voices, including its first generative voices for Hindi and Finnish. The production release keeps existing API calls and voice names intact. Teams already using generative Polly can receive different audio without changing their integration.

The [AWS announcement](https://aws.amazon.com/blogs/machine-learning/amazon-polly-delivers-a-major-upgrade-to-its-generative-engine/) describes better expressiveness, pronunciation and handling of structured text such as numbers and addresses. Those are AWS's claims, not independent measurements from this article. The release does not include a comparative quality score that would settle which voice works best for a particular application.

## New languages and regional availability

The expanded catalog contains 57 generative voices across 11 languages and 25 locales. The additions cover nine locales, including Indian English, Hindi, Finnish and Korean. AWS also added the engine to GovCloud US-West, bringing generative-engine availability to 11 regions.

Language support and hosting location are separate choices. A Hindi voice does not imply that the service is available in an India region: the announcement's region list does not include one. A team choosing a voice for Indian customers still needs to check where synthesis requests will run.

For a multilingual application, begin with the same representative script in each target language. Include names, abbreviations, dates and mixed-language text rather than testing only a clean introductory sentence. A fluent greeting tells you little about whether an appointment time or delivery address will be read correctly.

## Amazon Polly generative pricing

AWS says the upgrade does not introduce a new price or SKU. The [Polly pricing page](https://aws.amazon.com/polly/pricing/) lists the following pay-as-you-go rates per million characters:

| Engine | Listed price per million characters |
| --- | --- |
| Standard | $4 |
| Neural | $16 |
| Generative | $30 |
| Long-form | $100 |

At the listed generative rate, 100,000 billed characters cost $3 before any applicable credits or other charges. These are character prices, not token prices or prices per minute of audio. Confirm the applicable region's pricing before budgeting a deployment.

Polly allows generated speech to be cached and replayed without an additional synthesis charge. That makes the distinction between repeated and personalized audio important. A fixed introduction can be generated once; a new account-specific message requires new synthesis.

The practical cost comparison should therefore count newly generated text, not every time an audio file is played. An application with many repeated prompts can have a very different bill from one that narrates unique responses throughout a conversation.

## Existing APIs do not guarantee identical audio

The upgrade preserves the generative engine's existing voice identifiers and API contract. Standard and neural engines are unaffected. For an existing generative integration, a migration project may be unnecessary, but an audio acceptance test is still useful.

A sensible regression set should include the phrases whose delivery matters most: amounts, reference numbers, short instructions and pauses before important details. Compare the generated files with the application's expectations and keep approved static audio where consistency is important.

AWS's [generative-voice documentation](https://docs.aws.amazon.com/polly/latest/dg/generative-voices.html) explains that the engine produces speech incrementally. That supports responsive playback, but perceived responsiveness also depends on the application: buffering, network time and when playback begins can all affect what a listener experiences. A model-quality update does not measure that complete path for you.

## SSML support still needs checking

Speech Synthesis Markup Language, or SSML, lets an application annotate text with instructions such as pauses and pronunciation. Support varies by Polly engine. The [supported-tag table](https://docs.aws.amazon.com/polly/latest/dg/supportedtags.html) marks pauses and paragraph tags as available for generative voices, while emphasis is unavailable and prosody support is partial.

This matters when reusing a script built for another engine. Valid XML is not enough to establish that every intended delivery instruction is supported. Test the actual tags and voice combination rather than assuming that switching the engine preserves the performance.

The useful opportunity here is broader language coverage without a new API integration. The remaining work is to verify the words customers actually hear, because compatibility at the request level and consistency at the listening level are different acceptance tests.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn product updates and integration tradeoffs into clear technical stories that developers and founders read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).