Sustia

Project · 2026

Community voice dataset creation

An open-access pipeline for communities building voice datasets — recorded or digitized, segmented, transcribed locally, reviewed, exported — with Indigenous data sovereignty as the design center.

Status
in progress
Client
Open-access / community use; M1 work for Tauhokohoko
Stack
PythonuvWhisperMeta MMSJupyterLJSpeech
Tags
voice-dataindigenous-aipython

Problem

Communities want to preserve language and train voice models without sending audio to cloud APIs or losing control of consent and provenance.

Outcome

An end-to-end local-only pipeline — record/digitize → segment → SNR-check → transcribe (Whisper/MMS/manual) → review → LJSpeech export — built around CARE/OCAP governance, with consent and exclusion flags travelling with the data.

The repository started in 2020 as a tool to produce LJSpeech-format training data for TTS models. The 2026 reframe replaces that lens entirely: instead of how do we get training data? the question is how does a community maintain authority over its own voices?

That reframing changed the architecture. The pipeline is now end-to-end local — no cloud transcription, no third-party hosting of audio. Consent, provenance, and exclusion flags travel with the data through every step.

What the pipeline does

A community can enter at one of three pathways: record new audio with silence-based segmentation; bring existing recordings and segment manually; or augment a small consented dataset with speed/pitch/noise variations. All three converge on an SNR quality check, a language-routed transcription step (Whisper for major languages, Meta MMS for the 530+ languages MMS covers — Cree syllabics, Ojibwe, O’odham, and many more — manual for everything else), a transcript review pass, metadata entry, and LJSpeech export. The export step filters by consent tier; if a contributor withdraws, their recordings disappear from the next build without anyone rewriting upstream.

Governance bound to the data

Most of the rework was governance documentation, not code:

  • A plain-language community-agreement template covering ownership tiers, cultural protocols, and withdrawal rights.
  • An OCAP-aligned datasheet for the dataset as an institution, not just a file.
  • A “what not to digitize” decision framework for recordings that should not become training data.
  • A CARE-Principles file explaining how Collective Benefit, Authority to Control, Responsibility, and Ethics are enacted in this repo specifically.

The rules are versioned alongside the code. A community forks the repo and forks the governance with it.

What changed under the hood

  • Transcription moved local. The 2020 version used Google Cloud Speech-to-Text. The 2026 version routes by language to local Whisper, Meta MMS, or a manual workflow.
  • Hashed dependencies. uv with locked, hash-verified installs — no surprise wheel substitutions.
  • No cloud credentials at any step. Audio doesn’t leave community infrastructure.

Status

In active use. Companion to the chapter AI Techniques for Indigenous Cultural Expression in the forthcoming Envisioning Indigenous Methods in Digital Media and Ecologies, edited by Ashley Cordes. Current threads on Now. Repository: github.com/hollygrimm/voice-dataset-creation.