Amazon Polly Generative upgrade: time to move off Neural voices?

The Amazon Polly Generative engine has been upgraded in place, so existing generative voices now sound different with no code change, and 14 new voices arrive, including Hindi and Indian English. For teams still on Neural voices, the question is whether more natural speech is worth a higher price per character and the loss of speech marks.

What was released

The upgrade was announced by AWS on 22 October 2026 in Amazon Polly delivers a major upgrade to its Generative engine. AWS describes it as a drop-in upgrade for existing generative voices: voice names and API calls stay the same, and applications already calling Polly with the generative engine move to the new version without any change on their side. It is generally available now.

What actually changed

• More expressive speech. AWS says generative voices now adjust emotion and prosody from the input text without manual setup, and read numbers, currency, addresses and confirmation codes more accurately.
• New voices and languages. The release adds 14 voices across 9 locales, among them Kiara and Arjun for Indian English and Hindi, Kenneth and Nate for US English, and Suvi for Finnish. Hindi and Finnish are available on the Generative engine for the first time. The catalogue is now 57 generative voices in 11 languages and 25 locales.
• More Regions. The engine now runs in 11 AWS Regions, including AWS GovCloud (US-West), but each new voice is available only in some of them.
• Same price. AWS states there is no price increase, no new SKU and no integration change. The Amazon Polly pricing page still charges generative voices per character at close to double the Neural rate, with a generative free tier for new accounts.
• Streaming for voice agents. The post points to the bidirectional StartSpeechSynthesisStream API, which takes text in and sends audio out at the same time, for real-time uses such as voice agents.
• One gap remains. Speech marks, the word and sentence timing data some apps use for captions, highlighting or lip sync, are still not available on generative voices. Standard and Neural voices keep them and are not affected by this upgrade.

Should you move off Neural voices?

In CodeDTX's view, the decision comes down to three questions.

1. Does speech quality change your outcome? For a voice agent that reads back booking references, amounts or addresses, clearer numbers and natural tone can reduce repeat requests and transfers to people. For short, fixed IVR prompts, Neural may be good enough, and caching those prompts makes the engine choice matter less.
2. Can you afford the difference at your volume? Generative costs close to twice as much per character as Neural on the published rates. Work out your monthly characters from logs, not estimates, and see our guide to controlling AI agent running costs. Static prompts that you synthesise once and replay add no further cost.
3. Do you depend on speech marks? If your app highlights words as they are spoken, drives an avatar's lips or builds captions from Polly's timing data, stay on Neural or Standard for those flows until AWS adds speech marks to generative voices.

For teams serving callers in India, the new Indian English and Hindi voices are the main reason to look again: before this release, a Hindi voice agent on Polly had no generative option at all.

What to check if you already use generative voices

Your voice has already changed, so treat this like any silent model update:

• Replay your test calls and listen for how numbers, codes and names are read, not only whether the words are right. Our guide to what AI agent evals catch covers how to set this up, and the same steps from our Grok Voice 3.0 retest checklist apply here.
• Check your SSML. The generative engine supports SSML, but only the tags in the AWS list of supported SSML tags; prompts written for Neural may use tags it ignores.
• Confirm voices per Region. AWS recommends calling DescribeVoices in deployment checks, because a new voice may exist in one Region and not another.
• Watch production after the switch. Compare call completion, transfers and repeat calls before and after, following how to monitor an AI agent in production.

When not to switch

Stay where you are if your flows rely on speech marks, if your volume makes the higher rate hard to justify for little gain, or if a voice you need is not available in the Region your data must stay in. AWS has not said whether the previous generative version can be selected, so if you need fixed audio for regulated scripts, record and store approved prompts rather than synthesising them on every call.

Frequently asked questions

Do I need to change code to get the upgraded Polly voices?

No. AWS says the upgrade is in place: voice names and API calls are unchanged, and any application already calling Polly with the generative engine uses the new version automatically. Only teams on Neural or Standard voices need to change the engine setting to generative, and they should check that the voice they want exists in their Region first.

Did the Amazon Polly Generative upgrade change the price?

No. AWS states there is no price increase and no new SKU. Generative voices are still charged per character at a higher rate than Neural voices on the published Amazon Polly pricing page. Teams moving from Neural should estimate monthly characters from real logs before switching, and cache static prompts, which AWS says can be replayed at no extra cost.

Does Amazon Polly now have Hindi generative voices?

Yes. The upgrade adds Kiara and Arjun for Hindi and Indian English, and AWS says Hindi is available on the Generative engine for the first time. Availability of each new voice varies by AWS Region, so call DescribeVoices in the Region you deploy to before you build a Hindi voice agent around a specific voice name.

Can I use speech marks with Polly generative voices?

No. Speech marks, the timing data used for word highlighting, captions and lip sync, are still not available on generative voices after this upgrade. Standard and Neural voices continue to support them. If a flow depends on that timing data, keep it on Neural or Standard and use generative voices only where timing data is not needed.