Apple Silicon first · Phase 1, audio
Not one large neural network. Each engine picks whichever model and runtime wins on measured evidence, and nothing is promoted without a dated report behind it.
- Engines
- 8
- Complete
- 1
- Active
- 1
- With a champion
- 1
The spectrogram is synthetic and drawn in the browser. No audio is recorded on this page.
Measured, not claimed
What holds a champion title today
One engine has champions, per language. The rest have none, and the index says so in the same voice.
Audio / ASR / TTS
ASR English: whisper-large-v3-turbo-mlx (WER 1.36%). ASR Khmer: whisper-turbo-khmer-v9 (CER 0.148 [0.127,0.173]) (ADR-0013). TTS: none
The gate
A champion is promoted, never declared.
Every model on the index went through the same five steps. That is what makes the status column worth reading.
- 01
A candidate arrives with its own report
No model is integrated before a dated candidate report exists in the repository. The report is the ticket in.
- 02
The same test set, every time
Same utterances, same reference text. A number on a different set is a different experiment, not a comparison.
- 03
Same preprocessing, decoding and hardware
Feature geometry, normalisation, decoding settings and the machine are held fixed, or the difference measures the setup, not the model.
- 04
A paired comparison, with its uncertainty
Utterance by utterance against the current champion, bootstrapped. A win that is not significant is recorded as a tie.
- 05
Promoted, or kept as standby
The winner is promoted per language, the decision is written down, and the runner-up stays listed with what it is better at.
The index
Every engine, including the ones that do not exist yet
Projected out of docs/STATE.md, and held to it by a check that fails the build when the two disagree.
- 00Complete
Platform kernel
- 01Active
Audio / ASR / TTS
ASR English: whisper-large-v3-turbo-mlx (WER 1.36%). ASR Khmer: whisper-turbo-khmer-v9 (CER 0.148 [0.127,0.173]) (ADR-0013). TTS: none
- 02Not started
OCR / Document AI
- 03Not started
Language / LLM
- 04Not started
Computer vision
- 05Not started
Reinforcement learning
- 06Not started
Generative
- 07Not started
Multimodal / omni
Decisions on the record
Why the status column can be trusted
Sixteen decision records so far. Each says what was chosen, why, and what would reverse it. These are the ones that shaped what you read above.
- ADR 0003
Honest backend detection
Empirical probing of what the machine can run. No silent fallback.
- ADR 0004
Manifests in Git, payloads out
Every dataset, model, adapter and evaluation has a tracked manifest. The blobs stay outside.
- ADR 0008
Adopt proven work
Re-derive a method only when it is an oracle for a test or in the charter's study scope.
- ADR 0009
Whisper stays champion, just
Paired-bootstrap significant against Parakeet, and only just. Parakeet is the speed and memory optimum.
- ADR 0010
A standby champion
Qwen3-ASR is statistically tied with Whisper at twice the speed, and is the Khmer fine-tuning target.
- ADR 0012
Whisper cannot transcribe Khmer
Despite listing it. Khmer needs a fine-tune, so champions are per language.
- ADR 0013
A Khmer champion
whisper-turbo-khmer-v9, promoted through the same gate, with its confidence interval on the record.
- ADR 0016
Time-boxed, resumable fine-tuning
Whisper LoRA sessions with a ledger. No training run has started, and the page will say so until one has.
A champion is promoted, never declared
Same test set, preprocessing, decoding and hardware, or the comparison means nothing and the model does not move.
No model without a dated report
Nothing is integrated before its own candidate report exists in the repository, so every claim has a document behind it.
The kernel never imports an engine
One shared core, and engines that cannot reach into each other. A new engine is added, not threaded through.
The console is a shell, and says so
The interface is built ahead of the engines on purpose. It is wired to nothing yet, and it will tell you that rather than pretending.
Open the console