ARK-ASR-3B-GGUF

ARK-ASR-3B GGUF

GGUF conversions of Audio8/ARK-ASR-3B, made from source revision 1e28271b79edc97635783bea65abc89195a09ed3.

ARK-ASR combines a Whisper-style audio encoder, an MLP adapter, and a Qwen2.5 decoder. These files use GGUF architecture arkasr; they are not compatible with stock llama.cpp or whisper.cpp. They require the ARK-ASR implementation in the accompanying transcribe.cpp fork used by Handy.

Runtime

transcribe.cpp-arkasr-b838a2b.tar.gz is a source snapshot of the exact ARK-enabled runtime used for validation (commit b838a2b). SHA-256: 776c2a7a690dc21dc629aa79d03003a0a9f9a463f1e0a4f3a4cefc6ee59dbe97.

On macOS with Xcode command-line tools:

tar -xzf transcribe.cpp-arkasr-b838a2b.tar.gz
cd transcribe.cpp-arkasr
cmake -S . -B build -DTRANSCRIBE_BUILD_EXAMPLES=ON -DCMAKE_BUILD_TYPE=Release
cmake --build build --target transcribe-cli -j
build/bin/transcribe-cli --backend metal --model ../ark-asr-3b-Q8_0.gguf input.wav

The CLI expects WAV audio. The model frontend itself expects 16 kHz mono samples; the Handy desktop integration performs its normal recording conversion before inference.

Files

Local validation

The models were tested on an Apple M4 Max with the Metal backend. Six saved English recordings (6.6–79.2 seconds, 724 F16-reference words total) were compared against an F16 GGUF generated from the official BF16 checkpoint.

This is a small regression set, not a published WER benchmark. “Word disagreements” measures divergence from the F16 runtime output, not correctness against human transcripts.

An additional 491-second saved recording (1,413 F16-reference words) was used as a long-form check on the same M4 Max:

Q8_0 remains the recommended model: the smaller variants save only a few seconds on eight minutes of audio while diverging more from the F16 reference.

Long audio

The ARK runtime accepts arbitrary-duration audio. It processes independent windows of at most 30 seconds, chooses a low-energy boundary between 20 and 30 seconds when possible, overlaps windows by 2 seconds, and stitches the final transcript. This is offline transcription, not simulated live streaming.

License and attribution

The original model is Apache-2.0 licensed. See the official model card and AutoArk repository for architecture, training, and upstream attribution.