PsarAI-ASR-1B-GGUF

GGUF builds of nphearum/PsarAI-ASR-1B — speech recognition for Khmer, English and mixed Khmer–English — for audio.cpp.

File	Type	Size	Content CER	Use
psarai-asr-1b-f16.gguf	F16	1.93 GB	7.90%	reference quality
psarai-asr-1b-q8_0.gguf	Q8_0	1.04 GB	7.87%	lossless
psarai-asr-1b-q6_k.gguf	Q6_K	814 MB	7.90%	lossless
psarai-asr-1b-q5_k.gguf	Q5_K	689 MB	7.93%	lossless
psarai-asr-1b-q4_k.gguf	Q4_K	571 MB	8.03%	recommended: best size / quality
psarai-asr-1b-q3_k.gguf	Q3_K	445 MB	8.22%	small loss
psarai-asr-1b-q2_k.gguf	Q2_K	349 MB	10.59%	clear loss; only if memory is very tight

Content CER on 360 held-out clips; the transformers model scores 7.89% on the same clips (details below).

Each GGUF embeds its config.json, vocab.json and the audio.cpp model spec (family mms_asr).

These files need audio.cpp with the mms_asr family. It is not in released audio.cpp yet (proposed upstream). Until it is merged, build audio.cpp with audiocpp/mms_asr.patch from this repository. Stock audio.cpp, and the audio.cpp prebuilt shipped with Unsloth Studio, cannot load them. These are not llama.cpp GGUFs.

Build audio.cpp with mms_asr

The patch applies to upstream commit 50749561 (audio.cpp v0.9.2):

git clone https://github.com/0xShug0/audio.cpp.git && cd audio.cpp
git checkout 507495611b22dc4590e0e5609e05e8096384e33f
git apply --3way /path/to/mms_asr.patch
cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DENGINE_ENABLE_CUDA=ON \
  -DAUDIOCPP_MODEL_SET=custom -DAUDIOCPP_MODELS=mms_asr     # or the full model set
cmake --build build --target audiocpp_cli audiocpp_server -j

With CUDA 12.x, use a host compiler it supports (for example CC=gcc-13 CXX=g++-13 -DCMAKE_CUDA_HOST_COMPILER=g++-13). For Unsloth Studio, start it with UNSLOTH_AUDIO_CPP_PATH pointing to a folder that contains bin/audiocpp_server from this build.

Run

./build/bin/audiocpp_cli --task asr --family mms_asr \
  --model psarai-asr-1b-q4_k.gguf --backend cuda --audio speech.wav

text_output=good morning, sir. សូមស្វាគមន៍ការត្រឡប់មុកវិញ។ i've completed the system diagnostics, and everything is functioning normally. តើលោកចង់ឱ្យខ្ញុំចាប់ផ្តើមការងារអ្វីមុន។

Any sample rate or channel count is accepted (resampled to 16 kHz mono). Audio longer than ~32 s is decoded in 30 s windows with 2 s context (--session-option window_sec=..., context_sec=...). The output is the greedy CTC transcript; numbers may come out as digits.

Accuracy and speed

Compared with the transformers model (nphearum/PsarAI-ASR-1B, fp32, greedy) on 360 held-out clips (Khmer read speech, Khmer news, synthetic English and mixed speech), RTX 3090:

	Identical transcripts	Content CER	Khmer news (grkpp) CER	English CER	360 clips
transformers (reference)	—	7.89%	14.27%	0.06%
F16	336 / 360	7.90%	14.32%	0.06%	39 s
Q8_0	272 / 360	7.87%	14.29%	0.06%	27 s
Q6_K	243 / 360	7.90%	14.31%	0.06%	~30 s
Q5_K	185 / 360	7.93%	14.33%	0.09%	~30 s
Q4_K	143 / 360	8.03%	14.63%	0.06%	29 s
Q3_K	87 / 360	8.22%	14.55%	0.11%	30 s
Q2_K	38 / 360	10.59%	17.93%	0.43%	32 s

F16 differs from the reference by single characters only (25 character edits in total). Q8_0 to Q4_K change more transcripts but keep the error rate within 0.15 points; Q3_K loses a little, Q2_K clearly (it starts misspelling even English, e.g. twelfe). Peak VRAM on a 99 s recording: ~2.7 GB (F16), ~1.9 GB (Q8_0); both run at about 60× real time there.

SHA-256: F16 ffb58b35…a25, Q8_0 0b751e35…30b, Q6_K 6d6c13e6…c0e, Q5_K 574c9931…2db, Q4_K 43c46807…d48, Q3_K 90259f11…ed, Q2_K 77fc2769…3b (full hashes on each file's page). See the main model card for per-test-set accuracy and limitations (English tested on one synthetic voice only, lowercase output, missing digits ៤ ៦ ៨ ៩).

License

The model weights are CC-BY-NC-4.0 (non-commercial), following the base model facebook/mms-1b-all; see the main model card for data licenses and credits. audiocpp/mms_asr.patch modifies audio.cpp (Apache-2.0, Copyright 2026 ShugoAI LLC) and is provided under the same Apache-2.0 license.