Text-to-speech (TTS) is technology that converts written text into spoken audio. Modern neural TTS uses deep learning to produce speech with natural rhythm, stress and tone, a big improvement over the robotic voices of older systems. TTS powers screen readers, voice assistants, navigation apps, audiobooks and phone systems.
This page covers the TTS pipeline and controls. For voice cloning, dubbing and the wider field, see What is AI Voice Generation?
Key takeaways
- TTS turns text into speech in three broad stages: text processing, acoustic modelling and waveform generation.
- Neural TTS replaced older concatenative and rule-based methods.
- SSML lets you control pronunciation, pauses and emphasis.
- TTS is essential for accessibility.
- Pronunciation of names, abbreviations and numbers needs testing.
Text-to-speech explained in simple terms
Reading aloud looks easy, but involves many decisions. Is “Dr.” “doctor” or “drive”? Is “2026” a year or a quantity? Where should you pause? Which word should you stress in “I didn’t say he stole it”? A TTS system makes all these decisions, then produces a voice that says the result naturally.
How neural TTS works
- Text normalisation: expand numbers, dates, currencies and abbreviations. “₹2,500 on 3/4” becomes words appropriate for the language and region.
- Linguistic analysis: convert words to phonemes (speech sounds) and predict prosody: rhythm, pauses, pitch and emphasis.
- Acoustic model: a neural network predicts acoustic features over time, commonly a mel spectrogram, which describes how sound energy varies across frequencies.
- Vocoder: another neural network turns those features into an audio waveform.
Some newer models are end-to-end, going from text or tokens directly to audio, and some treat speech as tokens generated by a language model, which can capture expressive, conversational delivery.
Milestones
| Era | Approach | Result |
|---|---|---|
| Early systems | Rule-based formant synthesis | Understandable but robotic |
| 1990s–2010s | Concatenative: stitching recorded snippets | More natural, but choppy and inflexible |
| 2016 | DeepMind’s WaveNet neural vocoder | Big jump in naturalness |
| 2017 onwards | Tacotron and successors: neural text-to-spectrogram | Natural prosody, easier new voices |
| 2020s | End-to-end and token-based models | Expressive, multilingual, real-time speech |
Controlling speech with SSML
SSML (Speech Synthesis Markup Language) is a W3C standard supported by many TTS services. It lets you add tags to text to control:
- Pauses (for example, a half-second break).
- Emphasis on specific words.
- Pronunciation using phonetic spelling or aliases (useful for brand names).
- Speaking rate, pitch and volume.
- How numbers and dates are read.
A worked example: product name pronunciation
A company called “QueryFeedz” finds TTS reading it as “kwery-fedz” with odd stress. Using an SSML alias or phonetic tag, they specify it should be read as “query feeds”. They also check “Pvt. Ltd.” is read as “Private Limited” and that prices in rupees are voiced correctly. Testing a pronunciation list before launch avoids awkward errors across hundreds of audio files.
Where TTS is used
- Accessibility: screen readers for blind and low-vision users; reading support for dyslexia.
- Assistants and IVR: voice responses in apps, cars and phone systems.
- Content: audio versions of articles, courses and audiobooks.
- Navigation and alerts: turn-by-turn directions and announcements.
- Language learning: hearing correct pronunciation.
TTS vs speech-to-text
TTS turns text into speech. Speech-to-text does the opposite, turning speech into text. Voice assistants use both: speech-to-text to understand you, and TTS to reply.
Limitations
- Mispronounces rare names, jargon, mixed languages and acronyms.
- Long passages can sound monotonous or have odd emphasis.
- Emotional delivery is improving but not always appropriate to context.
- Voice quality varies by language and provider.
- Cloned or highly realistic voices raise consent and fraud concerns.