Text-to-Speech

Text-to-speech turns written text into audible speech.

Early systems stitched together recorded fragments, which sounded exactly like stitching.

Modern neural systems generate the audio, and the quality gap is enormous.

The usual pipeline predicts a spectrogram from text, then converts it to a waveform.

That second stage is the vocoder, and it is where much of the realism comes from.

Newer models generate audio tokens directly and decode them in one stage.

Prosody is the hard part: rhythm, stress and intonation.

The same words carry different meaning depending on emphasis.

Models infer that from the text, and long or unusual sentences expose the limits.

Homographs are a persistent problem, since read and read are spelled identically.

Voice cloning can copy a speaker from a very short sample, sometimes seconds.

That is useful for accessibility and obviously open to abuse.

Consent, watermarking and detection are live concerns rather than theoretical ones.

Latency matters for conversation, since a natural reply needs to start within a few hundred milliseconds.

Streaming synthesis starts speaking before the whole sentence is generated.