Views
No views yet
License: Apache 2.0 (inherited from source model). See the original model card for full terms.
| File | Size | Type |
|---|---|---|
cohere-transcribe-arabic-f16.gguf | 4.1 GB | F16 (reference precision) |
cohere-transcribe-arabic-q8_0.gguf | 2.4 GB | Q8_0 |
cohere-transcribe-arabic-q4_k.gguf | 1.5 GB | Q4_K |
cohere-transcribe-arabic-q4_k-imatrix.gguf | 1.5 GB | Q4_K + importance matrix (Arabic-calibrated) |
cohere-transcribe-arabic-ref.gguf | small | per-stage reference activations for crispasr-diff |
-imatrix build is
calibrated on CC0 Common Voice Arabic (fsicoli/common_voice_17_0, ar/dev) and
recovers quality at the 4-bit budget; prefer it over plain q4_k.transcribe.cpp reference runtime (~1.5 s, load excluded).1# Build CrispASR (see the repo for full instructions)
2git clone --recursive https://github.com/CrispStrobe/CrispASR
3cd CrispASR
4cmake -G Ninja -B build -DCMAKE_BUILD_TYPE=Release
5cmake --build build -j
6
7# Shorthand: auto-downloads the recommended imatrix GGUF and defaults
8# the language to Arabic — no -m / --hf-repo / -l needed.
9build/bin/crispasr --backend cohere-ar audio.wavcohere-ar is a CLI alias for the cohere backend: it routes to the same
runtime, resolves cohere-transcribe-arabic-q4_k-imatrix.gguf via -m auto,
and sets -l ar unless you pass an explicit -l (which always wins — useful
if you want to run this model with the LID pre-step or force another
language for testing). It's equivalent to:1build/bin/crispasr --backend cohere \
2 --hf-repo cstr/cohere-transcribe-arabic-07-2026-GGUF:cohere-transcribe-arabic-q4_k-imatrix.gguf \
3 audio.wav -l ar-m at a locally downloaded GGUF:build/bin/crispasr --backend cohere -m cohere-transcribe-arabic-q4_k.gguf audio.wav -l aren and ar onlyconfig.json lists exactly two: en, ar. That matters more
than it sounds, because nothing else can tell you. The finetune keeps the
base tokenizer, so all 183 ISO-639-1 <|xx|> tokens are present in the vocab —
<|de|>, <|ru|>, <|ja|> all decode without error. And Cohere Transcribe
answers a wrong language fluently rather than failing. On one 8 s Arabic
clip, -l ru added a hallucinated leading word, -l ja swapped the quotation
marks for brackets, and -l de changed the diacritics — all plausible, none
flagged.cohere: language 'de' is not supported by this model — using 'en' instead. Supported: en, arcohere_transcribe.supported_languages. For any Cohere
GGUF converted before that key existed, declare it at runtime instead:CRISPASR_COHERE_LANGS=en,ar build/bin/crispasr --backend cohere -m old.gguf audio.wav -l auto-l auto, CrispASR identifies the language by
probing this model itself — one short decode per candidate, no whisper-tiny
download — and can therefore only ever return en or ar:cohere[lid]: en len=64 agree=1.00 div=0.79 score=158 :: The city is located in the city of Jerry, a large city of Je
cohere[lid]: ar len=82 agree=1.00 div=1.00 score=328 :: العاصفة شبه الاستوائية "جيري" تغا
crispasr: LID -> language = 'ar' (cohere-probe, p=0.675)| Component | Details |
|---|---|
| Encoder | 48-layer FastConformer, d=1280, heads=8, head_dim=160, ffn=5120, conv_kernel=9 |
| Decoder | 8-layer causal Transformer, d=1024, heads=8, head_dim=128, ffn=4096 (ReLU), max_ctx=1024 |
| Vocab | 16,384 SentencePiece tokens |
| Audio | 16 kHz mono, 128 mel bins, n_fft=512, hop=160, win=400, 8× time subsampling |
| Parameters | ~2B |
decoder_start_token_id (▁,
13764) to the control-token prompt, matching the reference decoder_input_ids.
Mel frame count is floor(n/hop)+1 (NeMo FilterbankFeatures), giving T_enc = calc_length(T_mel); both are required for the cross-attention context to line up.crispasr-diff)cohere-transcribe-arabic-ref.gguf holds per-stage reference activations dumped
from the transformers model by tools/dump_reference.py. Reproduce the
per-layer parity check with:1build/bin/crispasr-diff cohere cohere-transcribe-arabic-f16.gguf \
2 cohere-transcribe-arabic-ref.gguf audio.wavCohereLabs.apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.