arabic-asr

Cohere Transcribe Arabic

Cohere Transcribe Arabic is an open source 2B-parameter Arabic automatic speech recognition model for speech-to-text transcription. It is optimized for Arabic, Arabic Dialects, English, and Arabic-English code-switched speech. Use it for Arabic ASR, Arabic audio transcription, dialectal Arabic speech recognition, and English speech-to-text. The model uses a Conformer encoder-decoder architecture and is supported natively in Transformers. Based on the Cohere Transcribe architecture.

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

<style> @scope { th, td { text-align: left; padding: 0.375rem 0.625rem; letter-spacing: 0; vertical-align: top; line-height: 133.3333%; border: 1px solid #e0e0e0; } ul { list-style-type: disc; margin: 0; padding-left: 1em; li { margin: 0.25rem 0 0; line-height: 133.3333%; } } } </style> <table> <tbody> <tr> <th>Name</th> <td><strong>cohere-transcribe-arabic-07-2026</strong></td> </tr> <tr> <th>Architecture</th> <td>conformer-based encoder-decoder</td> </tr> <tr> <th>Input</th> <td>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.</td> </tr> <tr> <th>Output</th> <td>transcribed text</td> </tr> <tr> <th>Model</th> <td>a large Conformer encoder extracts acoustic representations, followed by a lightweight Transformer decoder for token generation</td> </tr> <tr> <th>Training objective</th> <td>supervised cross-entropy on output tokens</td> </tr> <tr> <th>Languages</th> <td> <ul> <li>Arabic</li> <li>English</li> </ul> </td> </tr> <tr> <th>License</th> <td>Apache 2.0</td> </tr> </tbody> </table>

✨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)

<details> <summary><b>Long-form transcription</b></summary>

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)

</details> <!-- <details> <summary><b>Punctuation control</b></summary> Pass `punctuation=False` to obtain lower-cased output without punctuation marks. ```python inputs_pnc = processor(audio, sampling_rate=16000, return_tensors="pt", language="ar", punctuation=True) inputs_nopnc = processor(audio, sampling_rate=16000, return_tensors="pt", language="ar", punctuation=False) ``` By default, punctuation is enabled. </details> --> <details> <summary><b>English transcription</b></summary>

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)

</details>

vLLM Integration

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

<details> <summary><b>Run cohere-transcribe-arabic-07-2026 via vLLM</b></summary>

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"

</details>

Results

<details> <summary><b>Open Universal Arabic ASR Leaderboard (as of 07.07.2026)</b></summary> <style> table.simple { border-collapse: collapse; } table.simple th, table.simple td { text-align: left; padding: 0.375rem 0.625rem; min-width: 5em; vertical-align: top; line-height: 1.2; border-bottom: 1px solid rgba(127,127,127,0.35); } table.simple thead th { white-space: nowrap; } table.simple th:first-child { position: sticky; left: 0; background: inherit; z-index: 1; } table.simple .num { text-align: right; } table.simple .highlight-row > th, table.simple .highlight-row > td { background: rgba(127,127,127,0.12); } table.simple .highlight-cell { background: rgba(127,127,127,0.18); } table.simple .metric { display: block; white-space: nowrap; } table.simple .metric-cer { display: block; font-size: 0.85em; opacity: 0.65; white-space: nowrap; } table.simple thead .metric-hint { display: block; font-size: 0.75em; font-weight: normal; opacity: 0.6; margin-top: 0.15em; } </style> <div style="overflow-x: auto;"> <table class="simple text-web3-14 font-body"> <thead> <tr> <th>Model</th> <th class="num">Average<span class="metric-hint">WER · CER</span></th> <th class="num">SADA<span class="metric-hint">WER · CER</span></th> <th class="num">Common Voice<span class="metric-hint">WER · CER</span></th> <th class="num">MASC clean<span class="metric-hint">WER · CER</span></th> <th class="num">MASC noisy<span class="metric-hint">WER · CER</span></th> <th class="num">MGB-2<span class="metric-hint">WER · CER</span></th> <th class="num">Casablanca<span class="metric-hint">WER · CER</span></th> </tr> </thead> <tbody> <tr class="highlight-row"> <th><strong style="white-space:nowrap">Cohere Transcribe Arabic 07-2026</strong></th> <td class="num highlight-cell"><span class="metric"><strong>25.87</strong></span><span class="metric-cer"><strong>11.80</strong></span></td> <td class="num"><span class="metric"><strong>37.47</strong></span><span class="metric-cer"><strong>23.53</strong></span></td> <td class="num"><span class="metric"><strong>5.82</strong></span><span class="metric-cer"><strong>1.62</strong></span></td> <td class="num"><span class="metric">19.60</span><span class="metric-cer">6.45</span></td> <td class="num"><span class="metric">27.07</span><span class="metric-cer">10.13</span></td> <td class="num"><span class="metric">15.54</span><span class="metric-cer">8.40</span></td> <td class="num"><span class="metric"><strong>49.71</strong></span><span class="metric-cer"><strong>20.66</strong></span></td> </tr> <tr> <th style="font-weight:normal;">OmniASR LLM 7B</th> <td class="num"><span class="metric">28.32</span><span class="metric-cer">12.52</span></td> <td class="num"><span class="metric">41.61</span><span class="metric-cer">24.95</span></td> <td class="num"><span class="metric">8.75</span><span class="metric-cer">2.71</span></td> <td class="num"><span class="metric">19.69</span><span class="metric-cer">5.76</span></td> <td class="num"><span class="metric">29.29</span><span class="metric-cer">10.66</span></td> <td class="num"><span class="metric">14.13</span><span class="metric-cer">7.10</span></td> <td class="num"><span class="metric">56.46</span><span class="metric-cer">23.96</span></td> </tr> <tr> <th style="font-weight:normal;">OmniASR LLM 3B</th> <td class="num"><span class="metric">29.96</span><span class="metric-cer">13.77</span></td> <td class="num"><span class="metric">46.18</span><span class="metric-cer">27.27</span></td> <td class="num"><span class="metric">9.15</span><span class="metric-cer">2.80</span></td> <td class="num"><span class="metric">19.90</span><span class="metric-cer">6.13</span></td> <td class="num"><span class="metric">30.03</span><span class="metric-cer">11.27</span></td> <td class="num"><span class="metric">14.22</span><span class="metric-cer">7.06</span></td> <td class="num"><span class="metric">60.27</span><span class="metric-cer">28.06</span></td> </tr> <tr> <th style="font-weight:normal;">OmniASR LLM 1B</th> <td class="num"><span class="metric">29.96</span><span class="metric-cer">13.40</span></td> <td class="num"><span class="metric">43.84</span><span class="metric-cer">24.54</span></td> <td class="num"><span class="metric">9.55</span><span class="metric-cer">2.97</span></td> <td class="num"><span class="metric">20.03</span><span class="metric-cer">6.14</span></td> <td class="num"><span class="metric">30.26</span><span class="metric-cer">11.18</span></td> <td class="num"><span class="metric">15.34</span><span class="metric-cer">7.56</span></td> <td class="num"><span class="metric">60.68</span><span class="metric-cer">28.02</span></td> </tr> <tr> <th style="font-weight:normal;">Cohere Transcribe 03-2026</th> <td class="num"><span class="metric">30.67</span><span class="metric-cer">16.37</span></td> <td class="num"><span class="metric">60.11</span><span class="metric-cer">45.44</span></td> <td class="num"><span class="metric">8.17</span><span class="metric-cer">2.49</span></td> <td class="num"><span class="metric"><strong>8.66</strong></span><span class="metric-cer"><strong>2.97</strong></span></td> <td class="num"><span class="metric"><strong>19.01</strong></span><span class="metric-cer"><strong>7.71</strong></span></td> <td class="num"><span class="metric">25.33</span><span class="metric-cer">9.28</span></td> <td class="num"><span class="metric">62.71</span><span class="metric-cer">30.31</span></td> </tr> <tr> <th style="font-weight:normal;">Qwen3-Omni 30B</th> <td class="num"><span class="metric">30.71</span><span class="metric-cer">13.67</span></td> <td class="num"><span class="metric">44.82</span><span class="metric-cer">26.11</span></td> <td class="num"><span class="metric">11.46</span><span class="metric-cer">4.28</span></td> <td class="num"><span class="metric">21.47</span><span class="metric-cer">5.59</span></td> <td class="num"><span class="metric">30.85</span><span class="metric-cer">11.28</span></td> <td class="num"><span class="metric"><strong>13.09</strong></span><span class="metric-cer"><strong>6.20</strong></span></td> <td class="num"><span class="metric">62.55</span><span class="metric-cer">28.53</span></td> </tr> <tr> <th style="font-weight:normal;">NVIDIA Conformer-CTC (LM)</th> <td class="num"><span class="metric">32.91</span><span class="metric-cer">13.84</span></td> <td class="num"><span class="metric">44.52</span><span class="metric-cer">23.76</span></td> <td class="num"><span class="metric">8.80</span><span class="metric-cer">2.77</span></td> <td class="num"><span class="metric">23.74</span><span class="metric-cer">5.63</span></td> <td class="num"><span class="metric">34.29</span><span class="metric-cer">11.07</span></td> <td class="num"><span class="metric">17.20</span><span class="metric-cer">6.87</span></td> <td class="num"><span class="metric">68.90</span><span class="metric-cer">32.97</span></td> </tr> <tr> <th style="font-weight:normal;">OmniASR LLM 300M</th> <td class="num"><span class="metric">32.96</span><span class="metric-cer">14.84</span></td> <td class="num"><span class="metric">51.38</span><span class="metric-cer">29.10</span></td> <td class="num"><span class="metric">12.03</span><span class="metric-cer">4.04</span></td> <td class="num"><span class="metric">20.66</span><span class="metric-cer">6.22</span></td> <td class="num"><span class="metric">32.45</span><span class="metric-cer">12.23</span></td> <td class="num"><span class="metric">16.58</span><span class="metric-cer">7.86</span></td> <td class="num"><span class="metric">64.64</span><span class="metric-cer">29.61</span></td> </tr> <tr> <th style="font-weight:normal;">Gemma 4 E4B</th> <td class="num"><span class="metric">32.98</span><span class="metric-cer">13.71</span></td> <td class="num"><span class="metric">43.40</span><span class="metric-cer">20.96</span></td> <td class="num"><span class="metric">19.65</span><span class="metric-cer">7.48</span></td> <td class="num"><span class="metric">24.86</span><span class="metric-cer">7.76</span></td> <td class="num"><span class="metric">33.59</span><span class="metric-cer">12.25</span></td> <td class="num"><span class="metric">17.72</span><span class="metric-cer">8.67</span></td> <td class="num"><span class="metric">58.63</span><span class="metric-cer">25.11</span></td> </tr> <tr> <th style="font-weight:normal;">Qwen3-ASR 1.7B</th> <td class="num"><span class="metric">33.36</span><span class="metric-cer">12.33</span></td> <td class="num"><span class="metric">45.53</span><span class="metric-cer">19.90</span></td> <td class="num"><span class="metric">16.90</span><span class="metric-cer">5.06</span></td> <td class="num"><span class="metric">24.37</span><span class="metric-cer">5.72</span></td> <td class="num"><span class="metric">34.29</span><span class="metric-cer">10.84</span></td> <td class="num"><span class="metric">16.57</span><span class="metric-cer">6.25</span></td> <td class="num"><span class="metric">64.47</span><span class="metric-cer">26.23</span></td> </tr> <tr> <th style="font-weight:normal;">Voxtral-Small 24B</th> <td class="num"><span class="metric">34.47</span><span class="metric-cer">15.29</span></td> <td class="num"><span class="metric">50.82</span><span class="metric-cer">28.85</span></td> <td class="num"><span class="metric">15.25</span><span class="metric-cer">5.54</span></td> <td class="num"><span class="metric">23.96</span><span class="metric-cer">7.06</span></td> <td class="num"><span class="metric">34.43</span><span class="metric-cer">12.22</span></td> <td class="num"><span class="metric">16.03</span><span class="metric-cer">7.41</span></td> <td class="num"><span class="metric">66.30</span><span class="metric-cer">30.64</span></td> </tr> <tr> <th style="font-weight:normal;">NVIDIA Conformer-CTC (greedy)</th> <td class="num"><span class="metric">34.74</span><span class="metric-cer">13.37</span></td> <td class="num"><span class="metric">47.26</span><span class="metric-cer">22.54</span></td> <td class="num"><span class="metric">10.60</span><span class="metric-cer">3.05</span></td> <td class="num"><span class="metric">24.12</span><span class="metric-cer">5.63</span></td> <td class="num"><span class="metric">35.64</span><span class="metric-cer">11.02</span></td> <td class="num"><span class="metric">19.69</span><span class="metric-cer">7.46</span></td> <td class="num"><span class="metric">71.13</span><span class="metric-cer">30.50</span></td> </tr> <tr> <th style="font-weight:normal;">Gemma 4 E2B</th> <td class="num"><span class="metric">35.87</span><span class="metric-cer">15.34</span></td> <td class="num"><span class="metric">46.23</span><span class="metric-cer">23.47</span></td> <td class="num"><span class="metric">23.76</span><span class="metric-cer">9.13</span></td> <td class="num"><span class="metric">27.47</span><span class="metric-cer">8.99</span></td> <td class="num"><span class="metric">36.15</span><span class="metric-cer">13.93</span></td> <td class="num"><span class="metric">20.72</span><span class="metric-cer">10.15</span></td> <td class="num"><span class="metric">60.87</span><span class="metric-cer">26.35</span></td> </tr> <tr> <th style="font-weight:normal;">Whisper Large v3</th> <td class="num"><span class="metric">36.86</span><span class="metric-cer">17.21</span></td> <td class="num"><span class="metric">55.96</span><span class="metric-cer">34.62</span></td> <td class="num"><span class="metric">17.83</span><span class="metric-cer">5.74</span></td> <td class="num"><span class="metric">24.66</span><span class="metric-cer">7.24</span></td> <td class="num"><span class="metric">34.63</span><span class="metric-cer">12.89</span></td> <td class="num"><span class="metric">16.26</span><span class="metric-cer">7.74</span></td> <td class="num"><span class="metric">71.81</span><span class="metric-cer">35.04</span></td> </tr> </tbody> </table> </div>

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

</details>

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 Transcribe 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 [email protected] 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.

Citation

To cite this model please use the following bibtex:

@misc{shaun_cassini_2026,
	author       = { Shaun Cassini and Sebastian Vincent and Xiaolu Lu and Julian Mack and Dhruti Joshi and Pierre Richemond },
	title        = { cohere-transcribe-arabic-07-2026 (Revision 0a8193c) },
	year         = 2026,
	url          = { https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026 },
	doi          = { 10.57967/hf/9549 },
	publisher    = { Hugging Face }
}