AWS Upgrades Amazon Polly’s Generative Engine With 14 Voices and Two New Languages

Amazon Polly’s upgraded Generative engine is generally available

AWS announced on October 22, 2026 that an upgraded Amazon Polly Generative engine is generally available for production. The release adds 14 voices across nine locales and takes the Generative catalog to 57 voices across 11 languages and 25 locales. It is a managed text-to-speech product update, not a downloadable model or a new foundation-model API.

The upgrade is in place for applications already using Generative voices: AWS says voice names and API calls do not change. Standard and Neural voices are unaffected. That lowers migration work, but it also means a production voice can sound different without an application-code release; teams with tightly approved audio should treat the rollout as a model change and rerun acceptance samples.

The release adds 14 voices across nine locales

The announced additions are Kenneth and Nate for US English; Russell and Curtis for Australian English; Kiara and Arjun for Indian English; Tina for German; Carlos for US Spanish; Leo and Carolina for Brazilian Portuguese; Hyunduk for Korean; Kiara and Arjun for Hindi; and Suvi for Finnish. Kiara and Arjun appear in both Indian English and Hindi locale coverage, so “14 voices across nine locales” should not be read as 14 globally interchangeable identities.

Hindi and Finnish are generally available on the Generative engine for the first time. AWS provides launch-page samples for those languages and for English entity reading and contextual delivery. These are selected demonstrations, not a blinded multilingual evaluation; they do not establish uniform quality across accents, names, numbers, code-switching, long-form narration or domain vocabulary.

The API contract stays familiar and adds no new model identifier

The synchronous path remains SynthesizeSpeech with a VoiceId, Engine="generative" and an output format such as MP3. AWS’s launch example uses VoiceId="Kenneth" in us-east-1. DescribeVoices is the deployment-time check for whether a selected voice is available in a target region. AWS does not publish a separate versioned model ID for this upgrade.

For live voice applications, StartSpeechSynthesisStream accepts streaming text while returning synthesized audio. AWS positions that route for voice agents and interactive characters. The announcement does not publish first-audio, p50 or p95 latency, concurrency, quota or interruption measurements, so “low latency” remains a provider description rather than a workload-level service result.

The new voices span 11 AWS regions

AWS lists US East (N. Virginia), US West (Oregon), Europe (Frankfurt), Europe (Zurich), Europe (London), Asia Pacific (Sydney), Asia Pacific (Seoul), Asia Pacific (Singapore), Asia Pacific (Tokyo), Canada (Central) and AWS GovCloud (US-West). The GovCloud addition expands the Generative engine’s regional footprint to 11 regions.

Region presence does not by itself answer data residency, logging, retention, quota, failover or regulated-workload questions. Confirm current voice availability with DescribeVoices at deploy time, pin the intended region in clients, and test what the application does when a voice or region is unavailable rather than silently falling back to a different voice or engine.

Generative speech remains $30 per million characters

The public Polly pricing page lists Generative speech at $30 per one million characters outside the free tier. It lists 100,000 Generative characters per month for the first 12 months and says generated static prompts may be cached and replayed without an additional Polly charge. AWS says the upgrade introduces no cost increase and no new SKU.

Character pricing is not full voice-workflow cost. Add text generation, telephony or WebSocket transport, storage, observability, retries, regional transfer, moderation and human escalation. For dynamic agents, measure billed characters per completed task and cost per resolved interaction; a natural voice does not compensate for verbose model output or repeated synthesis after interruptions.

Speech marks remain unavailable for Generative voices

AWS says Generative voices continue to support supported SSML tags, but speech marks are not available. Products that need word or sentence timing, viseme synchronization or highlight alignment cannot assume the Generative engine is a drop-in replacement for Standard or Neural output even when the speech call itself is unchanged.

SSML support also needs per-tag verification against the current documentation. Test pronunciation aliases, breaks, dates, currency, addresses and multilingual text with the exact voice. Keep a text or caption path independent of synthesized audio, and do not infer accessibility conformance from the availability of more natural speech alone.

AWS’s quality claims need application-level listening tests

AWS describes the upgraded output as more natural, expressive and accurate, and says voices can reflect emotion and prosody from textual context without manual setup. The launch page demonstrates Tiffany, Kenneth, Hindi and Finnish samples and quotes one Finnish publisher saying Suvi sounded noticeably more natural in its testing. These are provider-selected evidence and a named customer testimonial, not independent comparative results.

Build a fixed evaluation set before approving the change: ordinary prose, product and person names, account identifiers, addresses, currencies, dates, acronyms, mixed-language passages, emotionally sensitive text and adversarial punctuation. Run old retained audio against new output where possible, use fluent native listeners for each locale, and separately score intelligibility, pronunciation, naturalness, inappropriate affect, clipping and time to first playable audio.

What voice and agent teams should do next

Adopt the upgrade for a bounded trial when an existing Generative voice is already approved or Hindi and Finnish close a real locale gap. Record the exact text, voice, locale, region, API route, output settings and retrieval date. Compare audio on representative scripts, confirm SSML behavior, monitor latency and errors, and preserve a rollback or alternative-engine path for customer-facing flows.

Constrain consequential uses until entity-reading and style tests pass. Wait when speech marks, a missing region or reproducible model versioning are required. Reject automatic use for high-impact spoken confirmations when a pronunciation or number error could authorize the wrong action without an independent on-screen or human confirmation. The release expands access; it does not remove the need to qualify each voice.

Disclosures

AccessAllGPT received no AWS account, Polly quota, audio, SDK access, credits, briefing, preview, review, payment or compensation for this article. AWS, Amazon Polly and the quoted customer did not sponsor, review or endorse it. AccessAllGPT did not test, score or rank any voice or engine. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Amazon Web Services or organizations cited. Publication-wide relationships are listed on the disclosures page.