Amazon Polly Generative Engine Upgrade: 14 New Voices, Hindi and Finnish, Same $30 Price

Amazon has upgraded the Amazon Polly Generative engine, its premium text-to-speech tier, and says the new version sounds more expressive, reads numbers and addresses more accurately, and keeps low latency. It adds 14 new voices across 9 locales, makes Hindi and Finnish generally available on the engine for the first time, and expands to AWS GovCloud (US-West). The catch for skeptics: the AWS announcement gives sample audio and feature claims, but no benchmark scores. This post separates what AWS states from what you should test yourself.

TL;DR

What AWS says improved

AWS names three quality areas, each demonstrated with sample audio in the post.

1. More natural, expressive delivery. The sample voice Tiffany is described as sounding "natural, expressive, and a little bit delightful."
2. Accurate entities. The demo has Kenneth read back an order confirmation with a tracking number, date and time, currency, address and a confirmation code. Contact-center IVR is the target: if the voice mangles a code, the caller escalates to a human.
3. Style from context. AWS says the voices reflect emotion and prosody "based on the text input, without manual setup." Its examples contrast a neutral "Your package arrived early." with an uncertain "So, um... your package arrived early. I think?" and a surprised "Wait, what? Your package arrived early?"

None of these are numerically measured in the post. Treat them as design goals backed by demos until you run your own tests.

New voices and languages

Hindi and Finnish are the strategic additions. Hindi opens a large market for voice agents and e-learning, and Finnish is a small-language case where good neural voices are scarce. AWS quotes Sanoma Media Finland as turning real-time news into audio, with over 10,000 articles per month using Polly's Finnish neural voice, and says the new Suvi voice improves on it per that customer's testing. That is a vendor-selected testimonial, not an independent evaluation.

Pricing: what is stated

AWS says the upgrade arrives at the same price with no cost increase, no new SKU and no integration changes, and that you can cache and replay static prompts at no additional cost. The Polly pricing page lists:

The Generative free tier is 100 thousand characters per month for the first 12 months. A quick calculation: a 1,000-word script is roughly 6,000 characters, so one million characters is about 160 such scripts, or about $0.19 per script at the Generative rate. Voice agents that speak every turn will use far more, so streaming volume is the number to model.

Building voice agents with it

AWS positions Polly in a cascaded pipeline: speech recognition, an LLM reply, then text-to-speech. The Bi-Di Streaming API lets you stream text from the model into Polly and receive audio simultaneously, which cuts the wait before the first sound. AWS lists the main benefits as low-latency playback, scale and cost efficiency.

A pragmatic evaluation plan:

1. Build a test set of 50 to 100 lines from your own domain: names, SKUs, phone numbers, currencies, addresses, dates, and code-mixed text if you serve Hindi or Indian English.
2. Compare old and new by regenerating the same lines. Since the upgrade is in place, save audio from before if you have it.
3. Measure time to first audio over streaming, not just final quality.
4. Check SSML behavior, which AWS says remains supported for fine control.
5. Have native speakers rate Hindi and Finnish output, because one demo sentence proves little.

The four Polly tiers, and where Generative sits

Polly has accumulated engines over the years, and the pricing page shows the ladder clearly: Standard at $4, Neural at $16, Long-Form at $100 and Generative at $30 per million characters. Long-Form costs more than Generative because it targets extended narration such as audiobooks and articles, while Generative is aimed at conversational and interactive speech. If you currently use Neural voices for an IVR or assistant, the upgrade makes Generative a candidate to evaluate: it costs about 1.9 times the Neural rate by the listed prices, which is our arithmetic, and the question is whether the quality gain justifies that on your traffic. Because the upgrade is in place, teams already on Generative get the improvement without a migration.

Use cases AWS highlights

AWS lists six scenarios, and each implies a different test.

• Contact centers. The emphasis is entity accuracy and low latency. Test with read-backs of tracking numbers, balances and addresses, the lines where a wrong digit costs a call.
• Cascaded voice agents. Text from an LLM is converted to speech and streamed over WebSockets or telephony. Test time to first audio and behavior when the LLM streams short fragments.
• Media, localization and dubbing. One pipeline across more markets, now including Hindi and Finnish. Test long scripts for consistent pacing.
• E-learning and training. Courseware that updates when the source text changes, with no re-recording. Test acronyms and product names with custom lexicons.
• Accessibility. Read-aloud and screen-reader style use. Test punctuation-heavy and table-like text.
• Interactive and transactional experiences. Games, kiosks and IoT devices. Test short utterances, which are where robotic prosody is most noticeable.

A simple cost model for a voice agent

Suppose an agent speaks about 400 characters per call turn, with 10 turns per call, so 4,000 characters per call. At $30 per million characters that is about $0.12 per call for speech output alone, before recognition, the LLM and telephony. At 100,000 calls per month, speech output is about $12,000. These are illustrative numbers from the published rate, not a quote. Caching helps: AWS says you can cache and replay synthesized static prompts at no additional cost, so greetings, menus and legal disclaimers should be generated once and stored, leaving only dynamic content to stream.

Migration checklist

1. Confirm your voices are Generative voices; Standard and Neural voices are not part of this upgrade.
2. Snapshot current audio for a regression set before and after, if you still have old outputs.
3. Re-test any SSML that adjusts rate, pitch or pauses, since changed prosody may interact with it.
4. Re-check downstream steps such as audio compression and silence detection, which can be sensitive to timing changes.
5. For Hindi or Finnish, add native-speaker review to your release gate.
6. For regulated work, verify the GovCloud (US-West) feature set in the AWS documentation before assuming parity with commercial Regions.

How it fits the voice AI market

Voice is crowded and moving fast. Recent posts on this blog cover alternatives and neighbors, each with different trade-offs:

• Open, tiny models: Gradium Phonon, a 31M on-device TTS.
• Real-time speech-to-speech: ElevenLabs Eleven Realtime and Grok Voice 3.0.
• Pricing claims to verify: HeyGen's voice TTS on OpenRouter.
• Agent SDK voice stacks: Mistral Voxtral 3.
• Transcription on the input side: Microsoft MAI-Transcribe-2.

Polly's pitch is not to win a leaderboard. It is that a managed AWS service with IAM, existing contact-center integrations and GovCloud availability can be adopted without a new vendor review. For regulated and public-sector teams, the GovCloud (US-West) listing is arguably the most consequential line in the announcement.

Trade-offs to weigh

• Cost versus open models. Running a small open TTS model yourself can be cheaper at scale but moves operations to you.
• Latency. AWS says low latency but gives no millisecond figures; measure from your region.
• Voice choice. 57 generative voices is a catalog, not unlimited cloning. If you need a custom brand voice, check whether Polly offers what you need before committing.
• Disclosure. Synthetic voices on phone lines are increasingly regulated; tell callers they are speaking with an AI where rules require it. See our coverage of AI voice in Parliament for an example of how voice cloning is being handled in public life.

What is not known

• Any independent mean-opinion-score, word-error or pronunciation-accuracy test.
• Latency numbers for the upgraded engine.
• The architecture or whether a new model underlies the upgrade; AWS calls it a "major upgrade" of the engine without details.
• Whether Regions beyond the 11 AWS lists (including GovCloud US-West) will get the new voices; AWS also notes speech marks are not available on generative voices.

Related reading

• Gradium Phonon open TTS
• ElevenLabs Eleven Realtime
• Grok Voice 3.0
• Mistral Voxtral 3
• MAI-Transcribe-2
• HeyGen voice TTS claim check

Details are as published by AWS on October 22, 2026; check the Polly documentation for current Region and voice availability.