Amazon Polly
A cloud service that converts text to natural-sounding speech, offering a neural speech engine and fine-grained speech control through SSML
Overview
Amazon Polly is a Text-to-Speech (TTS) service that converts text to natural-sounding speech in real time. Its neural TTS engine generates human-like natural speech, supporting over 30 languages including Japanese with dozens of voice options. In addition to fine-grained control of pauses, emphasis, and pronunciation through SSML (Speech Synthesis Markup Language), it offers speech style selection such as newscaster and conversational styles for different use cases.
Neural TTS Engine Mechanics and Speech Quality Evolution
Polly offers two engine types: standard and neural. The standard engine uses concatenative synthesis, joining recorded speech fragments to generate utterances. The neural engine, on the other hand, uses deep learning models to directly generate speech waveforms, resulting in more natural intonation and inflection with significantly reduced mechanical artifacts. The quality difference is particularly pronounced in long-text reading. The neural engine has further evolved into the Generative engine, with the latest Long-Form engine achieving natural prosody that considers context across paragraphs. Japanese neural voices include Kazuha and Tomoko, providing sufficient quality for business use. Azure Speech Service offers similar neural TTS, but Polly's advantages in practice are API simplicity and ease of integration with the AWS ecosystem. Audio data can be output in MP3, OGG, and PCM formats, allowing format selection based on use case.
Expressive Speech Control with SSML and Voice Styles
Simply converting text to speech can result in mispronunciation of proper nouns or unnatural pauses. SSML enables specifying pause duration with break tags, emphasis with emphasis tags, exact pronunciation with phoneme tags using IPA (International Phonetic Alphabet), and speed, pitch, and volume adjustment with prosody tags. For technical terminology, phoneme tags with IPA prevent misreadings. The newscaster style is optimized for news script reading, delivering speech in a calm broadcast-like tone. The conversational style is designed for chatbots and interactive applications, producing a more approachable tone. Speech synthesis books on Amazon systematically cover practical SSML usage patterns. The lexicon feature allows defining pronunciation for specific words and phrases at the account level, eliminating the need to write SSML each time.
Architecture Design for Real-Time Delivery and Batch Processing
Polly's SynthesizeSpeech API returns audio streams in real time, making it suitable for immediate playback in web and mobile applications. For large text volumes, the StartSpeechSynthesisTask API executes asynchronous batch processing, outputting generated audio files to S3. Batch processing handles up to 100,000 characters per task, supporting full book narration and podcast generation. A common architecture in practice involves EventBridge detecting article publication from a content management system, triggering Polly batch tasks via Lambda, and delivering generated audio through CloudFront. Pricing is usage-based on requested character count, with the neural engine approximately 4x the standard engine price, though the quality difference makes neural engine adoption recommended. The SpeechMark feature provides timestamps for each word in the audio, enabling automatic subtitle synchronization and karaoke-style highlight display.