Speech Synthesis

Speech synthesis, or text-to-speech (TTS), is the technology that generates artificial speech from a text. It is the inverse task of speech recognition: where the latter goes from audio to text, synthesis goes from text to audio.

From recorded fragments to neural networks

The technique has come a long way. Classic systems used concatenative synthesis, stitching together small fragments of pre-recorded human voice; later came parametric synthesis, which generated the sound's parameters with statistical models, gaining flexibility at the cost of some naturalness. The decisive leap came from deep learning: WaveNet (DeepMind, 2016) generated realistic audio sample by sample and drastically narrowed the gap with the human voice, and Tacotron (Google, 2017-2018) learned to produce the sound directly from text.

How it works today

A typical neural system works in two stages. First, an acoustic model converts the text into an intermediate representation of the sound, usually a spectrogram. Then, a vocoder transforms that spectrogram into the final audio waveform that we hear.

How quality is measured

The naturalness of a synthetic voice is evaluated mainly with the MOS (Mean Opinion Score): a panel of people rates the audio quality on a scale from 1 to 5, and their ratings are averaged. The best current systems approach the scores of recorded human speech.

Cloned voices and ethics

The most striking advance is voice cloning: systems capable of imitating a person's voice from very few seconds of recording. This opens up valuable uses—accessibility, dubbing, audiobooks—but also serious risks, such as the fraudulent impersonation of identity through fake voices (audio deepfakes). Consent and transparency have become the center of the legal and ethical debate around this technology.