Contribute your voice
Help the open voice community build speech recognition and voices for your language, by recording or reviewing speech.
Sentences are only one part of a voice assistant. Before Assist can match a sentence, something has to turn your voice into text. And when it answers, something has to turn text into a voice. Both are trained on recordings of people speaking, and for many languages there are not enough of them.
You can help, without writing a single line of YAML.
Speech to text transcribes what you say. To do that well, it needs to learn from recordings with a lot of variety:
- Many different speakers and accents.
- Different recording devices and quality levels, typically 16 kHz audio with 16-bit samples.
- Different places, like quiet and noisy rooms.
Contribute to Mozilla Common Voice. It is a free and open dataset of sentences spoken by people around the world. You can record sentences in your language, or help by listening to the recordings of others and validating them.
Text to speech turns text into a spoken voice. That needs the opposite of speech to text: consistency.
- A single speaker, or the same amount of recordings for each speaker.
- A high quality recording device, typically 48 kHz with 32-bit samples.
- A quiet, controlled place, like a soundproof booth.
Contribute to LibriVox, where volunteers record audiobooks of books in the public domain. Your recordings help the open voice community, and everyone gets free audiobooks out of it.
Curious which voices have come from the community so far? The voice datasets repository holds the public domain recordings used to train voices for Piper, the local text-to-speech engine of Home Assistant.