How the machines learn Thakk

Capture · AI in the loop from clip one

The corpus engine

Every recording that enters the vault is worked by machines so humans only do what humans must: voice-activity detection trims silence, quality scoring flags unusable audio at upload time, speaker diarization separates voices in elder conversations, and duplicate detection keeps the corpus honest.

Once the first ASR exists, it pre-transcribes the entire backlog so the paid transcription team corrects drafts instead of typing from silence — the single biggest cost lever in low-resource language work (3–5× faster per hour of audio). Active learning then closes the loop: the models tell us which kinds of speech the corpus lacks, and the prompt engine asks contributors for exactly that.

Models · adapt giants, own the result

The transfer-learning stack

Kodava Takk's neighbours — Kannada, Tulu, Malayalam — are already inside today's open multilingual models. We fine-tune rather than pre-train, so tens of hours of Thakk go a very long way and costs stay in lakhs, not crores:

• Hear — ASR: Whisper large-v3, Meta MMS / w2v-BERT, and AI4Bharat IndicASR checkpoints, LoRA-adapted on transcribed Thakk.
• Speak — TTS: VITS / Indic-TTS fine-tuned on consenting voice donors only; elder voice cloning solely with family + council consent.
• Translate — IndicTrans2 / NLLB for the Thakk–Kannada–English triangle, trained on sentence pairs produced during transcription.
• Converse — an open Indic LLM adapted on the Thakk text corpus, chained ASR → LLM → TTS into the Ainmane companion.

Evaluation · the community is the benchmark

Measured against our own ears

Every model release is scored two ways: word error rate on a held-out test set balanced across dialect (Mendele / Kiggat), age group, and genre — and a standing panel of fluent listeners who grade naturalness and correctness. A model that impresses a benchmark but sounds wrong to a Kodava grandmother does not ship.

Data · sovereignty by design

The guardianship pipeline

Consent tiers are enforced in code, not policy documents: an archive-only clip physically never enters a training run. Corpus releases are versioned like software, deposited to ELAR-standard archives and an Indian institutional mirror, and any model weights released publicly carry the Kodava data guardianship licence — usable for the community's benefit, never sellable. This is the Te Hiku Media doctrine, applied to Kodagu.

UNESCO · International Decade of Indigenous Languages (2022–2032) ↗

The UN system's global frame for indigenous language revitalization, including the Indigenous Language Data Commons Incubator (with The GovLab and Microsoft) for community-governed language data in the age of AI.

Why it matters to Thakk: Our guardianship model is designed to plug into this decade: community-led data commons, ethical AI norms, and international visibility for Kodava Takk.

Te Hiku Media · Papa Reo (Aotearoa NZ) ↗

The Māori-led language platform that trained te reo Māori speech recognition to ~92% accuracy and pioneered the Kaitiakitanga (guardianship) data licence — data governed by the community, never sold.

Why it matters to Thakk: The single most important proof that a small community can out-build big tech for its own language. Our licence, consent tiers, and council model follow theirs.

Indigenous Protocol & AI Working Group ↗

An international collective of indigenous technologists and scholars publishing position papers on how AI should be designed from indigenous epistemologies.

Why it matters to Thakk: The ethical playbook for every design decision where technology touches culture — from elder voice cloning to what 'consent' must mean.

Mila · First Languages AI Reality (FLAIR) ↗

The Québec AI institute's initiative building speech recognition for 200+ endangered North American indigenous languages, starting with the Wakashan family.

Why it matters to Thakk: State-of-the-art research on ASR when recordings are scarce — exactly our regime. A model for academic partnership done respectfully.

Mozilla Common Voice ↗

The world's largest open crowdsourced voice platform: 100+ languages built from 30-second clips, community validation loops, and public progress dashboards.

Why it matters to Thakk: Our Level 0 mechanics — tiny contribution units, validation by the community, visible counters — are the Common Voice playbook applied to Thakk.

AI4Bharat · IIT Madras ↗

IndicVoices (7,300+ hours across 22 Indian languages, 400+ districts), IndicASR, Indic-TTS, IndicTrans2 — open models and a district-by-district collection methodology, funded by Bhashini.

Why it matters to Thakk: Our nearest technical ancestors. Kodava Takk's Dravidian neighbours (Kannada, Tulu, Malayalam) live in these checkpoints — our fine-tunes start from them.

Project Vaani · IISc + ARTPARK + Google ↗

District-anchored open speech corpus for India — including recordings from Kodagu district — published openly on Hugging Face.

Why it matters to Thakk: Phase 0 of our roadmap audits Vaani's Kodagu audio for usable Kodava speech, and their field methodology informs our elder-session protocol.

Bhashini · Government of India ↗

MeitY's national language-AI mission funding open Indic datasets and models, with translation deployed at national scale.

Why it matters to Thakk: A funding and distribution channel: Kodava Thakk's corpus and models are designed to be Bhashini-compatible from day one.

Karya ↗

The Microsoft Research India spin-out that pays rural Indians dignified wages (with royalties) for voice data and annotation work in their own languages.

Why it matters to Thakk: Our paid transcription and validation layer follows Karya's model: corpus building as a modest livelihood program for Kodagu, not extraction.

Meta AI · MMS & No Language Left Behind ↗

Massively Multilingual Speech (ASR/TTS for 1,100+ languages) and NLLB (translation for 200 languages) — the open checkpoints UNESCO itself uses for its indigenous-language translator.

Why it matters to Thakk: Base models in our transfer-learning stack, alongside Whisper — we adapt giants, we never train from zero.

OpenAI Whisper (+ whisper.cpp) ↗

The open speech-recognition family that fine-tunes remarkably well on small corpora, with on-device inference via whisper.cpp.

Why it matters to Thakk: Our Phase 0 ASR demo is a Whisper fine-tune; whisper.cpp keeps future tools working offline in Kodagu's patchy connectivity.

ELDP & the Endangered Languages Archive (ELAR) ↗

Four decades of linguistic documentation standards: 48kHz WAV, time-aligned ELAN transcription, rich metadata, redundant archival deposit.

Why it matters to Thakk: AI models come and go; a properly archived corpus is forever. Everything we record is archived to these standards, mirrored beyond our own servers.

Google Arts & Culture · Woolaroo ↗

An open-source photo-translation tool for endangered languages using image recognition, live for languages from Māori to Yugambeh.

Why it matters to Thakk: A pattern for lightweight, joyful learning products our talking dictionary can borrow once the lexicon layer exists.