MMOOZZGG.com

Glossary

What is speech synthesis (text to speech)?

In short

Speech synthesis, also called text to speech (TTS), is a program that turns written text into a spoken voice.

AI basics · text-to-speech

MMOOZZGG.com editorsChecked: 1 min read

For a child

Remember the book with sounds: you press a button and hear the voice of a hero? A ready voice is recorded there. A computer can do more: you give it any text, even a brand-new one, and it reads it aloud by itself. It knows how each sound is pronounced, where to pause and when the voice should go up, as in a question. It sounds as if a real person were reading the book.

In plain words

The program gets a text and first prepares it for reading: it expands abbreviations (“St.” to “street”, “kg” to “kilograms”), writes numbers as words and marks the stress. Then the model decides how the phrase should sound: the pace, the pauses, the intonation. The last step builds a sound wave from all this. In the past, voices were glued together from recorded pieces, and they sounded broken. Neural networks speak smoothly, almost like a person. This is how navigation hints, stop announcements on public transport and articles for people with poor sight are voiced.

For business

Voice agents need speech synthesis to take calls, confirm bookings and remind clients about payment; it also voices training courses and videos. The quality of the voice decides whether a client hangs up in the first seconds. The price depends on the number of characters or minutes, the choice of a cloud or your own server, and the delay you need: in a live conversation the answer must sound almost at once. Use a real person’s voice only with their written consent. Check how the model reads your product names, addresses and amounts: stress errors in clients’ surnames annoy people more than a mechanical tone. The voice is chosen and tested on a pilot before launch.

For a specialist

Speech synthesis (text-to-speech) is usually split into three stages: text normalization and grapheme-to-phoneme conversion, an acoustic model that predicts a mel spectrogram, and a vocoder that builds the sound wave. WaveNet (van den Oord et al., 2016) generated raw audio autoregressively, one sample at a time, and in naturalness ratings beat parametric and concatenative systems for English and Chinese. Tacotron 2 (Shen et al., 2017) joined a recurrent seq2seq network that predicts mel spectrograms from characters with a modified WaveNet and got a mean MOS of 4.53 against 4.58 for a professional recording. Later work brought non-autoregressive and streaming models for low delay, and voice transfer from a short sample (zero-shot voice cloning). Quality is measured by the mean opinion score of listeners and by intelligibility, and intonation and pauses are marked up in SSML.

Example

A chain of dental clinics calls patients the day before a visit with a synthetic voice: it names the time and the doctor and asks the patient to confirm. The patient’s answer is recognized, and the schedule changes without an administrator.

Related terms

Questions and answers

Can you tell speech synthesis from a live voice? On short phrases, modern models sound very close to a human. They are usually given away by stress errors, odd pauses and a monotonous intonation in long text.
Can you voice a text with the voice of a specific person? Technically yes, since models need only a short voice sample. Do it only with that person’s consent: otherwise you violate their rights and open the way to fraud.

Sources

Where companies use it

Not sure whether your company needs this?

AI solutions map
Contact us