Text-to-Speech (TTS)

Text-to-speech (TTS) is the technology that converts written text into spoken audio. Modern TTS uses neural networks trained on recorded human speech to produce voices that sound natural — pitch, rhythm, pauses, and emotional inflection all mapped to the input text. A TTS system takes plain text as input (or structured markup like SSML) and outputs audio in formats like MP3 or WAV. TTS is used across accessibility tools (screen readers for visually impaired users), voice assistants (Alexa, Google Assistant, Siri), automated IVR systems in call centres, content creation (YouTube narration, audiobooks, podcasts), e-learning platforms, and language learning apps. The two major TTS approaches are concatenative synthesis (stitching together recorded speech fragments, now rare) and neural synthesis (generating waveforms end-to-end from text using deep learning). Neural TTS is the 2026 default — it handles prosody, intonation, and language-specific phonetic rules far better than the older concatenative approach.

How it works

A TTS pipeline typically runs in three stages: text normalisation (expanding abbreviations, numbers, dates, and symbols into spellable words), linguistic analysis (assigning phonemes, stress, and prosodic marks to each word), and waveform generation (producing the actual audio). Neural TTS models like Tacotron, FastSpeech, and their successors collapse these stages into a single deep learning model trained on paired text-and-audio datasets. Modern systems support voice cloning from short audio samples, style transfer between voices, and multi-language synthesis in a single model. TTS quality is typically measured by Mean Opinion Score (MOS) — a 1-5 rating of how natural the output sounds to native listeners.

Examples

Related terms

Learn more

Frequently Asked Questions