Text-to-Speech

What is text-to-speech?

Text-to-speech (TTS) is the artificial production of human speech from written text. A TTS system converts ordinary running text into audible speech; related systems also turn symbolic representations such as phonetic transcription into sound. TTS is a branch of speech synthesis and the counterpart to speech recognition (speech-to-text).

How does a TTS system work?

A classic TTS system has two parts. The front-end analyses the text: it normalises numbers, abbreviations and symbols into written-out words, assigns a phonetic transcription to each word, and divides the text into prosodic units such as phrases and sentences. The back-end — the synthesiser itself — converts this symbolic representation into sound, computing the target prosody, meaning the pitch contour and phoneme durations.

Which synthesis methods exist?

Several approaches have been used over time. Concatenative synthesis strings together recorded speech segments and sounds natural, but can produce audible glitches at the joins. Formant synthesis uses no human recordings at all, generating sound from an acoustic model, and stays intelligible even at high speed. Modern systems rely on deep learning: neural networks are trained on thousands of hours of recorded speech and generate the audio waveform directly from text. The first computer-based speech synthesis emerged as early as the late 1950s, but today's naturalness has only been reached with data-driven neural models.

How is TTS used for learning?

TTS makes content available through more than one channel. Study texts, summaries or flashcards can be read aloud and listened to on the go, which supports revision and helps people with visual impairment or reading difficulties. Because listening and reading engage different processing pathways, combining both channels can further support memorisation. Neural TTS is also the basis for turning personal documents into podcast-like audio formats. The voice quality of modern models is now high enough that trained listeners can barely tell the synthetic voice from a real recording.