GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Abdelrahman-M
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Abdelrahman-M
Model Tree
CohereLabs/cohere-transcribe-arabic-07-2026
Input Modalities
Output Modalities
Supported Functionality
| Source | Role |
|---|---|
| MASC (mined train) | ~73 h |
| MGB-2 train | ~50 h |
| SADA train | ~34 h |
| Common Voice 18 Arabic train | ~12 h |
| FLEURS ar_eg + MGB-3 Egyptian | ~13 h |
Total ~180 h. Official test splits of SADA, Common Voice, MGB-2, MASC, and Casablanca were not used for training.
Evaluated with the Tahqiq harness on the common / api slice (~1k utterances per set), greedy decoding, elm normalizer — same protocol as community leaderboard rows.
| Set | WER % | CER % |
|---|---|---|
| Casablanca | 40.59 | 15.50 |
| Common Voice 18 | 5.29 | 1.74 |
| MASC clean | 15.40 | 4.48 |
| MASC noisy | 23.38 | 7.41 |
| MGB-2 | 9.35 | 4.87 |
| SADA | 27.18 | 13.25 |
| Average | 20.20 | 7.88 |
| Avg WER | Avg CER | |
|---|---|---|
| Cohere Arabic 07-2026 | 23.59 | 10.81 |
| Arabic-Asr-cohere-v2 (this model) | 20.20 | 7.88 |
| Δ | −3.39 | −2.93 |
This model is supported natively in transformers. This is the recommended way to use it for offline inference. For online inference, see the vLLM integration example below.
bash
pip install transformers>=5.4.0 torch huggingface_hub soundfile librosa sentencepiece protobuf acceleratepython
from transformers import AutoProcessor, CohereAsrForConditionalGenerationfrom transformers.audio_utils import load_audio repo = "Abdelrahman-M/Arabic-Asr-cohere-v2"processor = AutoProcessor.from_pretrained(repo)model = CohereAsrForConditionalGeneration.from_pretrained(repo, device_map="auto") audio = load_audio("your_audio.wav", sampling_rate=16000)inputs = processor(audio, sampling_rate=16000, return_tensors="pt", language="ar")inputs.to(model.device, dtype=model.dtype) outputs = model.generate(**inputs, max_new_tokens=256)text = processor.decode(outputs, skip_special_tokens=True)print(text)For audio longer than the feature extractor's max_audio_clip_s, the feature extractor automatically splits the waveform into chunks. The processor reassembles the per-chunk transcriptions using the returned audio_chunk_index.
python
from transformers import AutoProcessor, CohereAsrForConditionalGenerationfrom transformers.audio_utils import load_audioimport time repo = "Abdelrahman-M/Arabic-Asr-cohere-v2"processor = AutoProcessor.from_pretrained(repo)model = CohereAsrForConditionalGeneration.from_pretrained(repo, device_map="auto") audio = load_audio("your_long_audio.wav", sampling_rate=16000)sr = 16000duration_s = len(audio) / sr inputs = processor(audio=audio, sampling_rate=sr, return_tensors="pt", language="ar")audio_chunk_index = inputs.get("audio_chunk_index")inputs.to(model.device, dtype=model.dtype) start = time.time()outputs = model.generate(**inputs, max_new_tokens=256)text = processor.decode( outputs, skip_special_tokens=True, audio_chunk_index=audio_chunk_index, language="ar")[0]elapsed = time.time() - startprint(f"Transcribed in {elapsed:.1f}s — RTFx: {duration_s / elapsed:.1f}")print(text)bash
vllm serve Abdelrahman-M/Arabic-Asr-cohere-v2 --trust-remote-codebash
curl -v -X POST http://localhost:8000/v1/audio/transcriptions \ -H "Authorization: Bearer $VLLM_API_KEY" \ -F "file=@$(realpath ${AUDIO_PATH})" \ -F "model=Abdelrahman-M/Arabic-Asr-cohere-v2"pip install transformers>=5.4.0 torch huggingface_hub soundfile librosa sentencepiece protobuf acceleratefrom transformers import AutoProcessor, CohereAsrForConditionalGenerationfrom transformers.audio_utils import load_audio repo = "Abdelrahman-M/Arabic-Asr-cohere-v2"processor = AutoProcessor.from_pretrained(repo)model = CohereAsrForConditionalGeneration.from_pretrained(repo, device_map="auto") audio = load_audio("your_audio.wav", sampling_rate=16000)inputs = processor(audio, sampling_rate=16000, return_tensors="pt", language="ar")inputs.to(model.device, dtype=model.dtype) outputs = model.generate(**inputs, max_new_tokens=256)text = processor.decode(outputs, skip_special_tokens=True)print(text)from transformers import AutoProcessor, CohereAsrForConditionalGenerationfrom transformers.audio_utils import load_audioimport time repo = "Abdelrahman-M/Arabic-Asr-cohere-v2"processor = AutoProcessor.from_pretrained(repo)model = CohereAsrForConditionalGeneration.from_pretrained(repo, device_map="auto") audio = load_audio("your_long_audio.wav", sampling_rate=16000)sr = 16000duration_s = len(audio) / sr inputs = processor(audio=audio, sampling_rate=sr, return_tensors="pt", language="ar")audio_chunk_index = inputs.get("audio_chunk_index")inputs.to(model.device, dtype=model.dtype) start = time.time()outputs = model.generate(**inputs, max_new_tokens=256)text = processor.decode( outputs, skip_special_tokens=True, audio_chunk_index=audio_chunk_index, language="ar")[0]elapsed = time.time() - startprint(f"Transcribed in {elapsed:.1f}s — RTFx: {duration_s / elapsed:.1f}")print(text)curl -v -X POST http://localhost:8000/v1/audio/transcriptions \ -H "Authorization: Bearer $VLLM_API_KEY" \ -F "file=@$(realpath ${AUDIO_PATH})" \ -F "model=Abdelrahman-M/Arabic-Asr-cohere-v2"