AWS 升級 Amazon Polly Generative 文字轉語音引擎，新增 14 種聲音

Amazon Web Services on October 22, 2026 released a generally available upgrade to the Amazon Polly Generative text-to-speech engine, adding 14 new voices across nine locales and bringing generative voices to Hindi and Finnish for the first time. In an AWS Machine Learning Blog post, the company described the updated engine as producing more expressive, more accurate speech output with low latency, now available in production.

The release arrives as a drop-in upgrade for existing generative voices: voice names and API calls are unchanged, so applications already using them keep working without modification. AWS describes Amazon Polly as a fully managed service that converts text into natural-sounding speech, used for contact centers, agentic applications built on large language models, media and entertainment, e-learning, accessibility, and other workloads that require reliable voice output at scale.

Speech Quality and Engine Design

AWS said the upgraded voices can adjust emotion and prosody to the input text without manual setup. The announcement demonstrates the behavior with three readings of the same sentence about a package arriving early — one neutral, one uncertain, and one surprised — and with sample audio. In one sample, the Tiffany voice greets a listener and introduces itself as one of Polly’s generative voices; in another, the Kenneth voice reads back an order confirmation, which AWS said accurately renders a tracking number, a delivery date and time window, a currency total, and a street address.

The Amazon Polly Developer Guide describes the Generative engine as the largest Amazon Polly text-to-speech model to date. It pairs a billion-parameter transformer, which converts raw text into speech codes, with a convolution-based decoder that turns those codes into waveforms in an incremental, streamable manner. The guide cautions that because the technology sits within the generative AI paradigm, updates to training data and the model can produce slight variations in how voices sound even when overall quality improves, a consideration for content synthesized in parts over a long period, such as a season of podcasts.

New Voices, Languages, and Regions

The 14 new voices, by locale, are 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). AWS published sample scripts for the Hindi voices Arjun and Kiara and the Finnish voice Suvi, and said Hindi and Finnish are generally available on the Generative engine for the first time.

With the additions, AWS said the Generative catalog now totals 57 voices across 11 languages and 25 locales. The 14 new voices are available in 11 AWS Regions: 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), with the GovCloud (US-West) addition expanding the engine’s footprint.

APIs, Pricing, and Documented Limitations

Existing integrations carry over unchanged. The same SynthesizeSpeech API call, with the engine parameter set to generative, reaches the upgraded voices, and AWS said no code changes are required. For real-time workloads, the bidirectional streaming API, StartSpeechSynthesisStream, lets an application stream text to Amazon Polly while receiving synthesized audio back at the same time, a pattern AWS described as well-suited to voice agents and bespoke characters in games.

Pricing stays the same. AWS said the upgrade delivers improved voice quality and accuracy at the existing price, with no cost increase, no new SKU, and no changes to current integrations, and that customers can cache and replay synthesized static prompts at no additional cost. The Generative engine also continues to support Speech Synthesis Markup Language tags for fine-grained control over output.

Several documented limitations remain in place. Speech marks are still unavailable on generative voices; standard and neural voices, which are unaffected by the upgrade, continue to support them. The developer guide notes that the newscaster speaking style is not supported in the Generative engine, that bidirectional streaming is offered only in certain regions, and that generative voices default to a 24 kHz sampling rate with MP3, OGG (Vorbis), and raw PCM output formats. The guide also describes a built-in emergency stop mechanism that halts speech rendering when the model shows potential to hallucinate, typically near the end of a sentence; AWS warns that the safeguard reduces but does not eliminate the risk, and that in edge cases the model can cut off part of a word or continue generating content beyond the stop threshold.

Availability and Early Customer Use

According to AWS, the quality improvements reach Amazon Connect and Genesys contact-center flows automatically. Developers can work with the voices through the AWS Management Console, the AWS Command Line Interface, or the AWS SDKs, using the polly:SynthesizeSpeech IAM permission, and can call DescribeVoices at deployment time to confirm that a selected generative voice is available in the region where an application runs.

Justus Reinikainen, a product manager at Sanoma Media Finland, said the company uses Amazon Polly to turn its real-time news articles into audio and currently publishes more than 10,000 articles per month with Polly’s Finnish neural voice. Reinikainen said that in the company’s testing, Suvi’s new generative voice was a clear improvement over the previous neural voice and sounded noticeably more natural for listeners.