Amazon Polly
Amazon Web Services Speech Synthesis service — Amazon Polly turns text into lifelike speech using deep learning, making it easy to create speech-enabled applications.
AI Practitioner focus
- Polly is the managed text-to-speech service. It accepts plain text or SSML and returns speech audio; current engines include standard, neural, long-form, and generative options where available.
- It does not translate language (Amazon Translate) or convert speech to text (Amazon Transcribe).
Key points
- Supports many different languages and Neural Text-to-Speech (NTTS) voices to build speech-enabled applications.
- Requires no setup costs; charges apply only for the amount of text converted.
- Offers caching and replay of generated speech in formats such as MP3.
- Engines — four TTS engines trade cost for naturalness: Standard (traditional, concatenative TTS), Neural (NTTS; deep-learning voices that sound noticeably more natural than standard), Long-form (tuned for longer narration — news articles, training materials, marketing videos — using text embeddings to produce more natural emphasis, pauses, and tone across a long passage), and Generative (Polly’s largest and most human-like, emotionally engaged, conversational engine, built on a billion-parameter transformer; positioned for use cases like a virtual assistant or trainer rather than long narration).
- SSML (Speech Synthesis Markup Language) — wrapping input in
<speak>tags unlocks control over pitch, speaking rate, volume, pauses/breaks, emphasis, phonetic pronunciation, whispering, breathing sounds, and the Newscaster speaking style, instead of relying on plain-text defaults. - Lexicons — custom pronunciation dictionaries (conforming to the W3C Pronunciation Lexicon Specification) that define how Polly should pronounce specific words, such as expanding an acronym (e.g., “W3C” → “World Wide Web Consortium”) or correcting a stylized spelling (“g3t sm4rt” → “get smart”).
- Speech Marks — metadata, not audio, returned alongside or instead of synthesized speech, marking where each sentence, word, viseme (mouth position), or SSML mark starts and ends in the audio stream; used for lip-sync animation and for highlighting/captioning words as they’re spoken. Available for standard, neural, and long-form voices; not currently available for generative voices.
- Fits a classic exam pipeline: Amazon Transcribe converts speech to text, Amazon Translate translates that text, and Amazon Polly converts the translated text back into speech.
Pricing
- Billed per character of input text, with separate per-engine rates: standard voices are the cheapest per million characters, neural voices cost more than standard, generative sits above neural, and long-form is the most expensive of the four engines per character.
- New accounts get a free-tier character allowance per engine (largest for standard, smallest for long-form/generative) for a limited introductory period — confirm current rates and free-tier limits on the AWS Polly pricing page, since they can change.