Parler-TTS v1: Hugging Face releases Mini and Large, voices controlled by a text description
In one sentence Hugging Face publishes Parler-TTS Mini v1 (880M) and Large v1 (about 2.2B), trained on 45,000 hours of audiobooks: the voice is described in words (gender, speed, pitch, noise). Open weights, data and training code, Apache 2.0 license.
With Parler-TTS you don't pick a voice from a menu: you describe it in words, in English, for example "an expressive female voice, moderate speed and pitch, very clean recording", and the model generates speech with those characteristics.
On August 8, 2024 Hugging Face releases two new checkpoints: Parler-TTS Mini (880 million parameters) and Parler-TTS Large (about 2.2 billion), trained on 45,000 hours of audiobooks. For a consistent voice across generations you can also name one of the 34 speakers used in training.
The project reproduces research by Stability AI and the University of Edinburgh, but releases everything: datasets, preprocessing, training code and weights, under the Apache 2.0 license.
Companies
HuggingFace
Tools
Parler TTS
Tags
Sources