Text-to-Speech
Generating natural-sounding speech from written text, including rhythm, intonation and, in some systems, a specific person's voice.
Also TTS · speech synthesis · voice cloning · vocoder
Reading aloud for the computer: it figures out how each word should sound, where to pause and which words to stress, then produces the voice.
Where it fits
+All relationships as a list
Is a kind of / part of
Siblings
Uses / built on
Used by
How it works
Text-to-speech runs speech recognition in reverse. A classic modern pipeline has two steps:
- Text to sound plan. An acoustic model turns text into a spectrogram-like plan: which sounds, how long each lasts, how pitch rises and falls.
- Plan to waveform. A vocoder neural network turns that plan into actual audio samples. WaveNet (2016) showed neural networks could do this far more naturally than older methods.
Newer systems treat audio like language: they split sound into discrete audio tokens and generate them with a Transformer, or use diffusion. Given a few seconds of someone’s voice, some can imitate it closely, which is useful for dubbing and accessibility, and a real risk for scams. Responsible systems require consent and add detectable watermarks.
Everyday use cases
- Case 01
Screen readers and accessibility
Reading websites and messages aloud for blind and low-vision users.
- Case 02
Navigation and assistants
Spoken directions and replies.
- Case 03
Audiobooks and dubbing
Narration and translated voice-overs.
Famous examples
Strengths, weaknesses and when to use it
Strengths
- Near-human naturalness in modern systems.
Weaknesses
- Voice cloning enables impersonation and fraud; consent and detection matter.
- Unusual names and mixed languages can be mispronounced.
When to use itWhenever information needs to be heard rather than read.
Compare with
Check yourself
- Q01
What does a neural vocoder do?
Show answer
A: Turns a spectrogram-like representation into an actual waveform