Amazon Polly’s new voices face a useful test: getting the address right

An automated phone system can sound convincing and still get the important part wrong: a parcel number, an order total or the street where a technician is supposed to arrive. Amazon’s latest Polly Generative upgrade targets that problem alongside more expressive speech. Announced on 20 October and explained in a 22 October technical post, the release also adds 14 voices across nine locales.[1][2]

AWS says the engine now offers 57 voices in 11 languages and 25 locales. Hindi and Finnish are new to the Generative engine, with Arjun and Kiara among the Hindi options and Suvi providing Finnish. Other additions cover existing language markets, including Australian English and Brazilian Portuguese.[1][2]

A confirmation code leaves little room for interpretation

A greeting can tolerate an extra pause or an unusual inflection. A booking code cannot tolerate a missing character. AWS demonstrates the update with text containing tracking details, a full address, a delivery time and an alphanumeric code. The examples make the intended improvement concrete, although they do not establish how often the engine makes such errors across real customer traffic.[1]

For an existing application, AWS describes the upgrade as a drop-in change: voice names and API calls remain the same. That reduces integration work, but developers who have approved a particular voice should still listen again. Polly’s documentation notes that generative-model updates can subtly change delivery, a concern when new narration must match earlier instalments of a course or publication.[1][3]

The same documentation lists a relevant limitation for accessible reading tools: generative voices do not supply speech marks, the timing information used to synchronise highlighted words with audio. Speech Synthesis Markup Language support also varies by tag. A more natural voice therefore does not automatically supply every feature a reading interface needs.[3]

The practical comparison is a representative script: local names, currency, abbreviations, quotations, addresses and codes, reviewed by fluent listeners in each target language. Counting corrections would show whether the upgrade saves production work. A voice that reads a greeting beautifully has passed one test; a voice that reliably tells someone where to go has passed a more consequential one.[1][3]

Report an error or suggest a correction