Text-to-Speech (TTS)

Text-to-speech, or TTS, synthesizes spoken audio from text. The system must produce not only recognizable sounds but also pronunciation, timing, pauses, and emphasis. It is an output-generation task, whereas speech recognition estimates text from input audio.

Written text leaves choices open

A support response containing a date, an abbreviation, and an order number needs an appropriate spoken rendering. An order number may need separate digits, while an amount should be read as a quantity. Text normalization resolves some of these choices before or within synthesis.

Tacotron 2 illustrates one architecture: it predicts a mel spectrogram, representing sound energy across frequency and time, and a vocoder converts that representation into a waveform. Other systems use different representations and generation methods, so this is an example rather than the definition of TTS.

A natural voice can still read the wrong thing

Suppose an approved message says a case requires review. Speech that changes a number or emphasizes a phrase misleadingly can fail even when the input text is correct. Listen for intelligibility, pronunciation, and preservation of task-critical details. A pleasant voice is not a check of the underlying policy.

A voice assistant may chain recognition, text generation, and TTS. Each stage has separate errors and latency. Producing speech in a particular style also does not by itself establish the speaker's identity or consent to imitate a voice; those are application-level requirements.

Reference: Shen et al.: Tacotron 2.