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:
1tar -xzf transcribe.cpp-arkasr-b838a2b.tar.gz
2cd transcribe.cpp-arkasr
3cmake -S . -B build -DTRANSCRIBE_BUILD_EXAMPLES=ON -DCMAKE_BUILD_TYPE=Release
4cmake --build build --target transcribe-cli -j
5build/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
| File | Size | SHA-256 | Notes |
|---|
ark-asr-3b-F16.gguf | 7,507,674,752 bytes | a9546be836351e17f616084f7ebc174570a076ebb7ee460bb551fc302af04176 | High-precision reference used as the quantization oracle |
ark-asr-3b-Q8_0.gguf | 4,289,201,792 bytes | b9ab32cfe7982eed5a596a601059e112d4aa5476e851852f944062a10fb25ee8 | Recommended accuracy-first build |
ark-asr-3b-Q6_K.gguf | 3,090,685,600 bytes | bf7e2fb7a50a9b5eed9b9e4eb36ae6498131c60abb5c2de3e4e4b3d61173ab9f | Balanced experimental build |
ark-asr-3b-Q5_K_M.gguf | 3,078,096,544 bytes | 1457567b5a2d0f9a943a41d78586e08b54aa79360a821bd1b2ce5d56137f8bdd | Experimental; conservative F16 tied embedding/output head |
ark-asr-3b-Q4_K_M.gguf | 2,307,317,408 bytes | 7d232eab9c87f35946d0370bed9ce8c7337b1da94a238ad5d7d71c47ed65da1f | Speed-first experimental build |
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.
| Quantization | Word disagreements vs F16 | Exact clips | Relative inference time |
|---|
| Q8_0 | 2 / 724 | 5 / 6 | baseline |
| Q6_K | 12 / 724 | 5 / 6 | about 9% faster overall |
| Q5_K_M | 16 / 724 | 4 / 6 | about 5% faster overall |
| Q4_K_M | 16 / 724 | 4 / 6 | about 17% faster overall |
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:
| Quantization | Inference time | Word disagreements vs F16 |
|---|
| Q8_0 | 28.16 s | 2 / 1,413 (0.14%) |
| Q6_K | 25.45 s | 14 / 1,413 (0.99%) |
| Q4_K_M | 22.96 s | 36 / 1,413 (2.55%) |
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.
