Text-to-Speech (TTS)
Technology that converts text data into speech. Foundation technology for screen readers and voice assistants.
Text-to-Speech (TTS) is a technology that converts text data into human speech. It is used across a wide range of applications including screen readers, voice assistants (Siri, Alexa, Google Assistant), car navigation systems, and e-book narration features. For users with visual impairments, TTS is an essential means of accessing web content.
TTS processing consists of three major stages. The first stage, text analysis, performs word segmentation through morphological analysis, reading estimation for numbers and abbreviations, and disambiguation of homophones. The second stage, prosody generation, determines accent, intonation, and pause positions. The third stage, speech synthesis, generates the actual audio waveform. Since the late 2010s, deep learning-based synthesis methods (WaveNet, Tacotron, VITS, etc.) have gone into production, giving smoother transitions in intonation than the earlier concatenative and statistical methods. Descriptions such as "indistinguishable from a human voice" are wording used in the explanatory text of individual services, and whether a listener can tell the difference depends on the voice and on the content of the text.
Web browsers provide TTS functionality through the Web Speech API's SpeechSynthesis interface. Implementation requires just a few lines of code: speechSynthesis.speak(new SpeechSynthesisUtterance('Text to read')). Cloud services such as Amazon Polly, Google Cloud Text-to-Speech, and Microsoft's Speech service (formerly Cognitive Services Speech) offer advanced control through SSML (Speech Synthesis Markup Language), enabling fine-tuning of reading speed, pitch, pauses, and volume.
Japanese TTS faces unique challenges with kanji reading disambiguation. The character "生" can be read as "nama," "sei," "shou," or "ikiru" depending on context. Personal and place names often lack dictionary entries, making custom dictionaries and ruby annotation data valuable for improving accuracy. Japanese also has word-specific accent patterns (flat, initial-high, mid-high, final-high), and for words missing from the dictionary the pattern has to be estimated, which makes natural prosody generation harder.
TTS and screen readers are closely related but serve different roles. TTS is the engine that converts text to speech, while a screen reader is software that interprets on-screen information and passes it to the TTS engine. To improve web content accessibility, it is important to use semantic HTML, set appropriate ARIA attributes, and provide alt text for images so screen readers can correctly interpret the structure.
From a character counting perspective, the character count of a script gives a first indication of how long it will take to read aloud. Figures of about 300-400 characters per minute for Japanese narration and about 150-180 words per minute for English are used as rules of thumb, and managing the length of podcast scripts and video narration by character count is a way of working built on those rules of thumb. Character count and duration are not proportional, however: what decides the length of a reading is the number of spoken beats, not the number of characters, so the more numerals, letters, and symbols a script contains, the further the estimate drifts from the measured time, and for tight timing it is safer to synthesize the text and measure the playback time.