Draft: reuse the ARK runtime for ARK-ASR-0.6B

docs/models/arkasr.md

@@ -0,0 +1,63 @@
+# ARK-ASR candidate port
+
+This candidate reuses the ARK implementation and converter published by
+harshav in [ARK-ASR-3B-GGUF](https://huggingface.co/harshav/ARK-ASR-3B-GGUF).
+The source archive `transcribe.cpp-arkasr-b838a2b.tar.gz` has SHA-256
+`776c2a7a690dc21dc629aa79d03003a0a9f9a463f1e0a4f3a4cefc6ee59dbe97`.
+The runtime additions are MIT licensed; original weights are Apache-2.0.
+The original model and inference implementation are maintained by
+[AutoArk](https://github.com/AutoArk/open-audio-opd).
+
+The source was adapted to the current device-selection API and gate/up packing
+status contract. Optional Q/K/V projection biases were added to the shared
+causal decoder, preserving null defaults for other families. The converter
+now records variant identity, supported languages and automatic language
+detection, accepts a pinned source revision, and checks variant dimensions.
+
+## Variants
+
+`ark-asr-0.6b` uses a 896-dimension Qwen2 decoder with 24 layers, 14 attention
+heads and 2 KV heads. `ark-asr-3b` uses the existing 2048-dimension decoder.
+Both have a Whisper-style encoder with partial interleaved RoPE and a merge-4
+MLP adapter. The 0.6B name describes the decoder, not the full checkpoint.
+
+## Conversion
+
+```sh
+uv run scripts/convert-arkasr-to-gguf.py \
+  --input Edge0/ARK-ASR-0.6B \
+  --revision 45776b56d58cdfb2e2eb632f7e110f38684633e0 \
+  --variant ark-asr-0.6b --outtype q8_0 \
+  --output ark-asr-0.6b-Q8_0.gguf
+```
+
+The reused converter's Q8 path retains the tied embedding/output matrix and
+convolutions in F16, and norms/biases/frontend tensors in F32. This is a
+candidate conversion path; moving quantization to the standard C++ policy is
+part of upstream acceptance work.
+
+## Runtime contract
+
+16 kHz mono PCM; offline automatic-language transcription; cancellation;
+no timestamps, translation, streaming or recognition hints. A language option
+does not steer this port. Long audio is processed in windows of at most 30
+seconds, with low-energy boundaries, overlap and transcript stitching.
+
+```sh
+build/bin/transcribe-cli --backend metal -m ark-asr-0.6b-Q8_0.gguf samples/jfk.wav
+build/bin/transcribe-cli --backend cpu -m ark-asr-0.6b-Q8_0.gguf samples/jfk.wav
+uv run scripts/verify-arkasr-reference.py --model /path/to/pinned/checkpoint \
+  --audio samples/jfk.wav --device mps
+```
+
+## Acceptance status
+
+This is a draft candidate for
+[issue 190](https://github.com/handy-computer/transcribe.cpp/issues/190).
+It is not a canonical published model family. The existing default test suite,
+real-model transcription comparisons and original-checkpoint oracle provide
+initial evidence. The intake signoff, standard golden tensor manifest/dumper,
+tolerance profile, family fixture/structural/ABI smokes, full numerical gates,
+WER dataset evaluation and canonical artifact publication remain required
+before upstream acceptance/publication. Transcript comparisons must not be
+reported as human-ground-truth WER.