VoicePreserve
Plan the recordings. Keep the language yours.
Guidance on documenting an endangered language, with consent and community ownership.
What VoicePreserve does
Live in the Hub chat. Answers say when a figure is illustrative, and nothing is sent, booked or decided for you.
Gives guidance on documenting a language. It records no audio, trains no models and stores no consent records.
- Gives guidance on documenting a language: planning recordings, asking for consent, and keeping ownership with the community
- Looks up a language in a seven-entry sample catalogue (code, region, vitality). Most languages are not in it
- Sketches how much recorded speech a speech-recognition or text-to-speech model would need. Its figures are illustrative
- It transcribes and speaks nothing: the speech-to-text and text-to-speech steps return placeholders
- Not connected: recording, consent records, model training, publishing, export and deletion. It records no audio and stores no consent
What is real, what is illustrative
12 tools are declared for VoicePreserve. Each is one of four kinds, and the Hub enforces the kind when the tool runs.
Illustrative only (2)
Returns demo output, marked as illustrative, with no confidence score or source attached. Not a real result.
- Assess corpus readiness
- Lookup language profile
Not connected (10)
Would act in the world (contact, send, file, store, publish), so it is switched off. It never runs, and nothing is sent or stored.
- Collect corpus with consent
- Endangered text to voice
- Endangered voice to text
- Escalate to human
- Export community data
- Publish open dataset
- Record community consent
- Review with community steward
- Revoke consent and cascade
- Train speech model
Start with a question
Example questions. Pick one and the chat opens with it ready to send, or bring your own.
- “Only our grandparents still speak our language fluently — how do we start recording them before it is too late?”Ask VoicePreserve
- “We have 40 hours of recordings of our elders — is that enough to train a speech model our school can use?”Ask VoicePreserve
- “An elder who recorded with us last year has changed her mind — how do we remove her voice from everything?”Ask VoicePreserve
Where it runs
The Hub runs on the web. Other channels, and the languages and regions listed, record what the agent is designed for.
- Channels
- Web (live)
Voice, API (planned, not yet available) - Languages (designed)
- multilingual, community-led
- Designed for
- Global
- Organisations the design looks to
- Mozilla Common Voice, CLEAR Global, Endangered Languages ProjectNo affiliation or endorsement. These organisations publish guidance the design draws on, or are ones we hope to work with. No partner system is connected.
More than one goal at once?
The coordinator can pass one conversation to several agents and bring their answers back as one.
Open the coordinator