What is a neural voice, and how do you choose one?

What is a neural voice?

A neural voice is a synthetic voice generated by a neural network that has learned from recordings how a person speaks. The model itself predicts the melody, rhythm and tone of each sentence.

The result is smoother and less monotonous than the voices of the past, especially on long sentences. The step-by-step process is described in how text-to-speech works.

How does it differ from concatenative and formant voices?

A concatenative voice stitches together small pieces of recordings cut in advance; a formant voice computes the sound from rules about the shape of vowels and consonants. The first sounds human but choppy at the joins, the second stays intelligible but clearly robotic.

A neural voice neither stitches nor follows hand-written rules: it generates the sound in one go. On the other hand, it can make more surprising mistakes, for example on a rare word or an intonation.

Where do Readotte's voices come from?

Readotte's 10 voices come from three open models, run on our servers in Europe. In English: Alba, Peter and Eve (Pocket TTS, by Kyutai), Heart with an American accent and George with a British accent (Kokoro).

In French: Caro, Michael and Estelle (Pocket TTS), Léa and Hugo (Supertonic 3, by Supertone). For more on English reading, see English text-to-speech.

What do the licenses of these models say?

Each model is published under its own license: Pocket TTS under CC BY 4.0, Supertonic 3 under OpenRAIL-M, Kokoro under Apache 2.0. These licenses set what a service like Readotte is allowed to do with the model, such as running it on its servers, with certain conditions like crediting the author or respecting usage restrictions.

They do not say what you, as a user, may do with audio obtained through a service: the service's own terms decide that. The difference is explained in text-to-speech and commercial use.

Why do some French voices keep an accent?

Some of Readotte's French voices are derived from English-speaking voices, and they can keep a slight Québécois intonation. The model has learned French, but part of the original voice's melody remains.

It is not a reading error: the words are pronounced in French. If that intonation bothers you, try another French voice; each one has its own tone.

How do you choose the voice that suits you?

Play each voice's sample on the Voices page, then have it read a few minutes of a real text, the one you plan to listen to. A voice that sounds pleasant on one sentence can be tiring over an hour, and the reverse is also true.

You can then fine-tune it: pitch by ±4 semitones, warmth and clarity by ±6 dB, and expressiveness for the Pocket TTS voices. These settings apply to all your documents in the same language.

• For a course or a report: a clear, not very expressive voice that holds up at a high speed.
• For a story: a more expressive voice, listened to at 1× or 1.25×.
• For background listening: raise the warmth and lower the clarity a little if the voice sounds harsh.