Amazon Polly upgrades its Generative voices: 14 new, same price

Amazon Web Services announced on 22 October 2026 a major upgrade to the Generative engine in Amazon Polly, its managed text-to-speech service. In a post on the AWS Machine Learning blog, AWS says the upgrade makes speech more expressive and more accurate at low latency, and that it is generally available in production.

What AWS says changed

AWS lists 14 new voices across 9 locales: US English (Kenneth, Nate), Australian English (Russell, Curtis), Indian English (Kiara, Arjun), German (Tina), US Spanish (Carlos), Brazilian Portuguese (Leo, Carolina), Korean (Hyunduk), Hindi (Kiara, Arjun) and Finnish (Suvi). The catalogue totals 57 generative voices across 11 languages and 25 locales. Hindi and Finnish are generally available on the Generative engine for the first time, and the engine now runs in AWS GovCloud (US-West), for 11 AWS Regions in all.

On quality, AWS claims more natural delivery and more accurate reading of numbers, currency and addresses. Its sample audio has a voice reading back an order confirmation with a tracking number, delivery date, dollar total and confirmation code. It also says the voices adjust emotion and prosody from the text alone, without manual setup. AWS publishes no listening-test scores, error rates or comparison against other speech vendors, so these remain AWS's description.

What it costs and what you change

AWS says existing generative voices upgrade in place: voice names and API calls are unchanged, and there is "no cost increase, no new SKU, and no changes to your existing integration." New applications use the same SynthesizeSpeech call with the engine set to generative. For real-time work, the Bi-Di Streaming API (StartSpeechSynthesisStream) lets an application stream text in and receive audio back at the same time, which AWS positions for voice agents and game characters. AWS also says synthesized static prompts can be cached and replayed at no additional cost.

The post names contact centres, cascaded voice agents (a language model's reply text converted to speech and streamed over WebSockets or telephony), localisation, e-learning, accessibility and kiosks as uses. For a contact centre, AWS argues that correct read-back of an order number or address helps a caller resolve the issue in self-service instead of escalating to a person. That is a plausible mechanism but AWS gives no escalation figure.

A customer view, and what is missing

AWS quotes Justus Reinikainen, product manager at Sanoma Media Finland, who says the publisher turns its real-time news articles into audio with Polly and publishes over 10,000 articles a month with Polly's Finnish neural voice. He says that, based on Sanoma's testing, the new generative Suvi voice is a clear improvement and sounds noticeably more natural. This is a vendor-chosen testimonial about one language, with no test method or numbers.

The post does not state per-character pricing, latency figures, or whether the older neural voices change. It also does not say whether an in-place upgrade can alter the audio of already-approved scripts, which matters if you have recorded compliance wording or cached prompts. Those are open questions, not findings.

What to do

• If you already use Polly generative voices, replay your ten most important scripts (order read-backs, addresses, amounts) and compare them before and after, since the audio changes without a code change.
• If you run Hindi or Finnish voice flows on neural voices, test the new generative voices on your own content, and compare cost per thousand characters on the Amazon Polly pricing page.
• For voice agents, measure time to first audio with the streaming API on your own network path before committing to it.