Akuruއަކުރު AI
ދިވެހިOpen the app

Progress

Akuru builds on more than five months of Dhivehi research and model development. This is what exists now, and how it measures.

Scoreboard

Speech to text, word error

Starting point71.6%
Akuru today52.3%
71.6%Base55%v0.252.3%v1.0
↓ lower is better internal

Base Whisper turbo vs akuru-raalhu-P-1.0, 1.97 h of conversational Dhivehi (lower is better)

Speech to text, character error

Starting point31.1%
Akuru today23.5%
31.1%Base27.9%v0.123%v0.223.5%v1.0
↓ lower is better internal

Same test (lower is better)

English → Dhivehi translation (chrF)

Starting point50.4
Akuru today66
50.4Base63.9v1.062.6v1.166v1.2
↑ higher is better internal

Base MADLAD-400 vs akuru-bathheli-T-1.2, 300 held-out pairs (higher is better)

Dhivehi → English translation (chrF)

Starting point46.9
Akuru today57.4
46.9Base57.5v1.056.6v1.157.4v1.2
↑ higher is better internal

Same test (higher is better)

Thaana OCR, character error

Starting pointNo working model
Akuru today0.31%
2.42%v1.10.31%v2.0
↓ lower is better internal

akuru-loa-T-2.0 on 32 pages of scans and screenshots

Dhivehi chat exam, correctness (0–5)

Starting point1.46–1.97
Akuru today2.28
Frontier 4.91.61v1.02.28v1.1
↑ higher is better

akuru-dhoni-T-1.1 vs four other open setups, 200 questions; a frontier model scores 4.90, which is the gap we are closing

Milestones

  1. Research and local model work

    • Planning, Dhivehi data research and local model experiments
    • Groundwork for the speech, translation and OCR models

  2. Platform build

    • Speech lab, Dhivehi text normaliser and GPU training setup
    • Web app on Cloudflare Workers with chat, voice, translation and image reading
    • First Akuru speech model (Whisper fine-tune)

  3. Accounts and own models in the app

    • Accounts, 100 free credits a month and bank-transfer payments with automatic receipt checks
    • Dhivehi-first interface, landing page, custom domain and SEO
    • Voice chat, feedback loop (likes, corrections, recordings) and the first Dhoni chat models
    • Translator v1: chrF 50.4 → 63.9

  4. Cloud, OCR and the model register

    • Serving moved to serverless GPUs that scale to zero; training moved to A100s
    • Thaana OCR trained from scratch
    • Dhoni exam: 200 Dhivehi questions; Dhoni T-1.1 wins among open models
    • Speech model turbo v1: word error 71.6% → 52.3%
    • Model register and naming: Dhoni, Bathheli, Raalhu, Beru, Loa, Odi
    • App runs only Akuru models

  5. OCR v2

    • akuru-loa-T-2.0: page exam character error 2.42% → 0.31%, reads small screen text and Latin-font dates
    • This website

Where we are honest about gaps

  • Chat is beta: Dhoni scores 2.28/5 on correctness against 4.90 for a frontier model. Grounding in Dhivehi sources and a direct Dhivehi model are the next steps.
  • Speech to text still makes many mistakes on casual speech; the target is under 15% character error.
  • Text to speech uses open voices today; our own voice needs licensed studio recordings.
  • The app is in testing with a small group of users.