TTS (Text-to-Speech) — AI Glossary

Reference details

Related terms

• ASR (Automatic Speech Recognition)
• Multimodal AI

How it works

1. Text is first normalised, which expands numbers, dates, currency and abbreviations into the words they are spoken as — the stage that decides whether a figure is read as a year or a quantity.
2. Normalised text is converted to phonemes, either by a pronunciation dictionary or by a learned grapheme-to-phoneme model for words the dictionary lacks.
3. An acoustic model predicts prosody over that phoneme sequence: the duration of each phone, the pitch contour, and the energy, which together carry stress and phrasing.
4. Those predictions are rendered into an intermediate acoustic representation, commonly a mel spectrogram or a sequence of discrete audio tokens.
5. A vocoder converts that representation into a waveform, and this is the stage that determines whether the result sounds clean or artefacted rather than whether it sounds natural.
6. For interactive use the pipeline runs per phrase so audio begins playing before the full utterance is synthesised, which is why phrase boundaries rather than sentence boundaries set the latency floor.

Example: Where synthetic speech gives itself away, priced per day

Not tone — prosody on unfamiliar strings. Addresses, product codes, initialisms and mixed-language names are where pacing and stress go wrong, because the model has to guess how something is said. Evaluate on your actual content, not on the vendor's demo sentences. The rate is worth establishing rather than assuming, because it determines whether a pronunciation lexicon is worth maintaining. Suppose 3% of utterances contain an address, a product code or a surname outside the model's common distribution, and 40% of those produce a stress or pacing error a listener notices. That is 0.03 x 0.40 = 1.2% of utterances with an audible defect, and at 50,000 utterances a day, 600 of them. A lexicon covering the few hundred terms that actually recur in your catalogue removes most of that, which is a bounded piece of work against a recurring and countable defect.

Commonly confused with

Voice cloning is not the same as voice synthesis. Synthesis generates speech from text in some voice; cloning reproduces a specific person's voice, which carries consent and likeness obligations that ordinary TTS does not — a distinction that is legal before it is technical, since the same model does both and which one you are doing depends on whose voice the reference audio came from. A second confusion is between naturalness and intelligibility: the published scores measure the first, while whether a listener correctly heard a postcode is the second.

When to use it

Reach for it when:

• Content must be consumed hands-free or eyes-free
• Accessibility requires an audio channel alongside the written content
• Latency budgets allow streaming synthesis in a voice interface
• Where a pronunciation lexicon can be maintained for the terms that recur

Reach for something else when:

• Cloning a voice without documented consent
• Precise pronunciation matters and you cannot supply a lexicon
• The text is dense reference material better read than heard
• Judged on a naturalness score alone, which does not measure whether an identifier was heard correctly

Referenced by

This term is used by the following ModelRefs references:

• speech pipeline stack
• voice realtime stack

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to TTS (Text-to-Speech) — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path