Skip to content
Thrivant.AI
SDG 10: Reduced InequalitiesTier 2Live in the Hub · its figures are illustrative; this page describes the design

VoicePreserve

Languages are falling silent. Some are recorded first.

Of the world’s thousands of living languages, commercial AI serves only a small share. VoicePreserve is designed to help build speech technology for the rest — corpora collected consent-first, models owned by the communities who speak them, revocation that actually revokes. The archive is not the point; the speakers are. Today it has recorded no audio, trained no models and holds no consent records.

Channels (designed)
Web todayPlanned: API · Voice
Languages (designed)
multilingual · community-led
Designed for
Global
Tools
12 declared8 not connected · 4 illustrative
The speaker

Naserian’s grandchildren answer in Swahili

An illustrative portrait, from the agent’s own system prompt, not a real person: Naserian is sixty-seven, Maasai, from Loita Hills in Narok County, Kenya. She speaks Maa fluently, Swahili roughly, English not at all. Her grandchildren are forgetting Maa, because the tablet, the voice note and the radio call-in all expect Swahili or English.

VoicePreserve is being built to meet her — an archivist that would answer to the people in its recordings.

The wall

Eight frames, one of them empty

What the archive is designed to hold, at its smallest unit: a greeting, in the orthography its community chose, beside the consent that would make keeping it legitimate. Nothing on this wall has been recorded — the waveforms are drawn. Rest a pointer on a frame and its line traces itself again.

Specimen — the listening wall · an illustration, nothing recorded
  • Sopa

    “hello”

    Maa mas

    Loita Hills, Kenya

    Not recorded. A greeting, shown to illustrate what a consented recording would look like.

  • ᐅᓪᓛᒃᑯᑦ

    ullaakkut — “good morning”

    Inuktitut iku

    Nunavut, Canada

    Not recorded. A greeting, shown to illustrate what a consented recording would look like.

  • ᎣᏏᏲ

    osiyo — “hello”

    Cherokee chr

    Oklahoma & North Carolina, US

    Not recorded. A greeting, shown to illustrate what a consented recording would look like.

  • ⴰⵣⵓⵍ

    azul — “hello”

    Tamazight tzm

    Atlas Mountains, Morocco

    Not recorded. A greeting, shown to illustrate what a consented recording would look like.

  • イランカラㇷ゚テ

    irankarapte — “let me touch your heart, gently”

    Ainu ain

    Hokkaidō, Japan

    Not recorded. A greeting, shown to illustrate what a consented recording would look like.

  • A language not yet recorded.

    This is the work.

  • Allillanchu

    “are you well?”

    Quechua (Cusco) quz

    Andes, Peru

    Not recorded. A greeting, shown to illustrate what a consented recording would look like.

  • Bures

    “hello”

    Northern Sámi sme

    Sápmi — Norway, Sweden & Finland

    Not recorded. A greeting, shown to illustrate what a consented recording would look like.

No recordings exist yet. The greetings and orthographies are real; the waveforms are drawn, and no speaker’s voice is reproduced on this page. No community named here has recorded anything with us or agreed to anything. The design ties every clip to a consent record signed by a named community authority, governed by the CARE principles for indigenous data, with a cascade delete when the record is revoked. None of that is connected yet.

The design

Consent is a record, not a mood

The design starts from a contract, not a caption. Nothing would be collected, trained or published without a Community Stewardship Consent Record, signed by a named human authority — not the AI, not the project: a person, with a role, from the community. Consent is scope-bound — recognition does not licence synthesis — and excluded sacred topics are refused, under the CARE principles of the data-sovereignty-care-principles skill.

None of this is connected yet. Today VoicePreserve records no audio and stores no consent records, and whatever you type in the Hub is sent to and recorded by our AI platform provider.

In the design, revocation triggers a cascade:

  1. First

    Corpus rows deleted

    Every sample bound to the revoked record leaves the store.

  2. Then

    Mirrors withdrawn

    Published datasets carrying the samples are pulled.

  3. Then

    Models retrained or withdrawn

    Anything trained on the data is rebuilt without it.

  4. Last

    An audit, in the community’s language

    What was deleted, where, and when.

“A community deserves a transparent agent more than an impressive one.”
— VoicePreserve system prompt
The approach

Five moves, always in order

The designed sequence is the same for every brief from the Hub coordinator — and the first step is never skipped. In the Hub today VoicePreserve only gives guidance on documenting a language; collection, training and release are not connected.

  1. Resolve the language first

    Maa, Quechua and Karen each map to several ISO 639-3 codes; the iso-639-3-and-glottocode-lookup skill fixes the canonical code and Glottocode first.

  2. Grade the urgency

    Nine factors, six grades — the unesco-language-vitality-framework skill; three willing elders of a critically endangered language mean capture this season.

  3. Collect on Common Voice lines

    Short sentences, clips under ten seconds, two community validators per clip — the mozilla-common-voice-methodology skill.

  4. Train at the corpus’s true size

    Under ten validated hours means forced alignment, not neural ASR — the low-resource-asr-tts-playbook skill matches model to data.

  5. Review before release

    A named community panel can veto any release; every model ships with a card.

Field glossary

The words the work is done in

Four terms recur throughout the design.

CSCR
Community Stewardship Consent Record — the signed, scope-bound, revocable consent a community grants, naming signatory, licence, modalities, excluded topics.
Vitality grade
UNESCO’s six-level scale, 0 (extinct) to 5 (safe): do children still learn the language at home?
TK label
A Traditional Knowledge label from Local Contexts, layered over the licence to carry community terms a licence cannot.
Model card
Every released model’s public record: consent records, review panel, per-dialect evaluation, failure modes, cascade-delete policy.
The refusals

What it will never do

Six of the agent’s rules begin with never. Its tools are written to refuse without a consent record; the collecting, training and publishing they guard are not connected yet, so today these are the rules of a design.

  • Never collect, transcribe, synthesise, train or publish without an active consent record.
  • Never record ceremonial, sacred or excluded-topic content, even when the speaker offers.
  • Never name a speaker: metadata stops at language code, dialect and a coarse age band.
  • Never use caste, clan, gender, sexuality or religion as conditioning signals for a model.
  • Never publish without an approved steward decision — “we will retroactively review” is not consent.
  • Never ship a model without a model card; black-box models reproduce the harm it exists to undo.

And when a speaker shows distress, capture is designed to halt; in the planned deployment, a human steward is asked to take over.

“The right answer is sometimes ‘we will not build this’. Record refusals; honour them.”
— VoicePreserve system prompt
Public standards it follows

Whose methods it follows

No affiliation or endorsement. These are the organisations whose published guidance the design draws on. No partner system is connected.

The methodology is borrowed from the published work of the institutions that proved it: Mozilla Common Voice, CLEAR Global and its Gamayun pipeline, the Endangered Languages Project, Glottolog, and Local Contexts. The agent works from a small built-in catalogue of languages, not from any of their systems.

Try it yourself

Ask VoicePreserve

“Only our grandparents still speak our language fluently — how do we start recording them before it is too late?”
An example question. Bring this one, or your own. VoicePreserve says when a figure is illustrative, and nothing is sent, booked or decided for you.
Ask in the Hub— opens VoicePreserve in the Agent for Good Hub in a new tabFree to use. Opens in a new tab. The chat may be temporarily unavailable.