CohereLabs/cohere-transcribe-arabic-07-2026

Cohere Transcribe Arabic

Cohere Transcribe Arabic is an open source release of a 2B parameter dedicated audio-in, text-out automatic speech recognition (ASR) model. The optimized for Arabic and English, with a focus on Arabic dialect performance and Arabic-English code-switching. Based on the Cohere Transcribe architecture.

Developed by: Cohere and Cohere Labs. Point of Contact: Cohere Labs.

Name	cohere-transcribe-arabic-07-2026
Architecture	conformer-based encoder-decoder
Input	audio waveform → log-Mel spectrogram. Audio is automatically resampled to 16kHz if necessary during preprocessing. Similarly, multi-channel (stereo) inputs are averaged to produce a single channel signal.
Output	transcribed text
Model	a large Conformer encoder extracts acoustic representations, followed by a lightweight Transformer decoder for token generation
Training objective	supervised cross-entropy on output tokens
Languages	Arabic English
License	Apache 2.0

✨Try the Cohere Transcribe Arabic demo✨

Usage

Cohere Transcribe Arabic is supported natively in transformers. This is the recommended way to use the model for offline inference. For online inference, see the vLLM integration example below.

pip install transformers>=5.4.0 torch huggingface_hub soundfile librosa sentencepiece protobuf accelerate

Quick Start 🤗

Transcribe any audio file in a few lines:

from transformers import AutoProcessor, CohereAsrForConditionalGeneration
from transformers.audio_utils import load_audio
from huggingface_hub import hf_hub_download

processor = AutoProcessor.from_pretrained("CohereLabs/cohere-transcribe-arabic-07-2026")
model = CohereAsrForConditionalGeneration.from_pretrained("CohereLabs/cohere-transcribe-arabic-07-2026", device_map="auto")

# Example: transcribe Arabic audio
audio_file = "your_audio.wav"
audio = load_audio(audio_file, 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)

Long-form transcription

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.

from transformers import AutoProcessor, CohereAsrForConditionalGeneration
import time

processor = AutoProcessor.from_pretrained("CohereLabs/cohere-transcribe-arabic-07-2026")
model = CohereAsrForConditionalGeneration.from_pretrained("CohereLabs/cohere-transcribe-arabic-07-2026", device_map="auto")

audio = load_audio("your_long_audio.wav", sampling_rate=16000)
sr = 16000
duration_s = len(audio) / sr
print(f"Audio duration: {duration_s / 60:.1f} minutes")

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() - start
rtfx = duration_s / elapsed
print(f"Transcribed in {elapsed:.1f}s — RTFx: {rtfx:.1f}")
print(text)

English transcription

The model also supports English. Specify language="en":

inputs = processor(audio, sampling_rate=16000, return_tensors="pt", language="en")
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)

vLLM Integration

For production serving we recommend running via vLLM following the instructions below.

Run cohere-transcribe-arabic-07-2026 via vLLM

First install vLLM (refer to vLLM installation instructions):

uv venv --python 3.12 --seed
source .venv/bin/activate

uv pip install -U vllm==0.19.0 --torch-backend=auto
uv pip install vllm[audio]
uv pip install librosa

Start vLLM server

vllm serve CohereLabs/cohere-transcribe-arabic-07-2026 --trust-remote-code

Send request

curl -v -X POST http://localhost:8000/v1/audio/transcriptions \
 -H "Authorization: Bearer $VLLM_API_KEY" \
-F "file=@$(realpath ${AUDIO_PATH})" \
-F "model=CohereLabs/cohere-transcribe-arabic-07-2026"

Results

Open Universal Arabic ASR Leaderboard (as of 07.07.2026)

Model	AverageWER · CER	SADAWER · CER	Common VoiceWER · CER	MASC cleanWER · CER	MASC noisyWER · CER	MGB-2WER · CER	CasablancaWER · CER
Cohere Transcribe Arabic 07-2026	25.8711.80	37.4723.53	5.821.62	19.606.45	27.0710.13	15.548.40	49.7120.66
OmniASR LLM 7B	28.3212.52	41.6124.95	8.752.71	19.695.76	29.2910.66	14.137.10	56.4623.96
OmniASR LLM 3B	29.9613.77	46.1827.27	9.152.80	19.906.13	30.0311.27	14.227.06	60.2728.06
OmniASR LLM 1B	29.9613.40	43.8424.54	9.552.97	20.036.14	30.2611.18	15.347.56	60.6828.02
Cohere Transcribe 03-2026	30.6716.37	60.1145.44	8.172.49	8.662.97	19.017.71	25.339.28	62.7130.31
Qwen3-Omni 30B	30.7113.67	44.8226.11	11.464.28	21.475.59	30.8511.28	13.096.20	62.5528.53
NVIDIA Conformer-CTC (LM)	32.9113.84	44.5223.76	8.802.77	23.745.63	34.2911.07	17.206.87	68.9032.97
OmniASR LLM 300M	32.9614.84	51.3829.10	12.034.04	20.666.22	32.4512.23	16.587.86	64.6429.61
Gemma 4 E4B	32.9813.71	43.4020.96	19.657.48	24.867.76	33.5912.25	17.728.67	58.6325.11
Qwen3-ASR 1.7B	33.3612.33	45.5319.90	16.905.06	24.375.72	34.2910.84	16.576.25	64.4726.23
Voxtral-Small 24B	34.4715.29	50.8228.85	15.255.54	23.967.06	34.4312.22	16.037.41	66.3030.64
NVIDIA Conformer-CTC (greedy)	34.7413.37	47.2622.54	10.603.05	24.125.63	35.6411.02	19.697.46	71.1330.50
Gemma 4 E2B	35.8715.34	46.2323.47	23.769.13	27.478.99	36.1513.93	20.7210.15	60.8726.35
Whisper Large v3	36.8617.21	55.9634.62	17.835.74	24.667.24	34.6312.89	16.267.74	71.8135.04

Link to the live leaderboard: Open Universal Arabic ASR Leaderboard.

Resources

For more details and results:

• Technical blog post contains WERs and other quality metrics.
• Announcement blog post for more information about the model.
• The Open Universal Arabic ASR Leaderboard.

Strengths and Limitations

Strengths

Cohere Transcribe Arabic demonstrates strong transcription accuracy for Arabic and English. As a dedicated speech recognition model, it benefits from efficient inference via the Conformer encoder-decoder architecture.

Limitations

• Single language. The model performs best when remaining in-distribution of a single, pre-specified language. It does not feature explicit, automatic language detection and exhibits inconsistent performance on code-switched audio.

• Timestamps/Speaker diarization. The model does not feature either of these.

• Silence. Like most AED speech models, Cohere Transcibe Arabic is eager to transcribe, even non-speech sounds. The model benefits from prepending a noise gate or VAD (voice activity detection) model in order to prevent low-volume, floor noise from turning into hallucinations.

Model Card Contact

For errors or additional questions about details in this model card, contact labs@cohere.com or raise an issue.

Terms of Use: We hope that the release of this model will make community-based research efforts into Arabic speech more accessible. This model is governed by an Apache 2.0 license.