CohereLabs/cohere-transcribe-arabic-07-2026 · Hugging Face

来源：https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026 Cohere Transcribe Arabic 是一个开源发布的 2B 参数专用音频输入、文本输出自动语音识别（ASR）模型。该模型针对阿拉伯语和英语进行了优化，尤其注重阿拉伯语方言性能和阿拉伯语-英语代码切换。基于 Cohere Transcribe (https://huggingface.co/CohereLabs/cohere-transcribe-03-2026)架构构建。

开发方：Cohere (https://cohere.com/) 与 Cohere Labs (https://cohere.com/research)。联系人：Cohere Labs (https://cohere.com/research)。

名称cohere-transcribe-arabic-07-2026架构基于 Conformer 的编码器-解码器输入音频波形 → 对数梅尔频谱图。预处理过程中，如有需要，音频会自动重采样至 16kHz。同样，多通道（立体声）输入会平均为单通道信号。输出转录文字模型一个大型 Conformer 编码器提取声学表征，后接一个轻量 Transformer 解码器用于生成 token训练目标输出 token 的有监督交叉熵语言- 阿拉伯语

• 英语

许可协议Apache 2.0✨试用 Cohere Transcribe Arabic 演示 (https://huggingface.co/spaces/CohereLabs/cohere-transcribe-arabic-07-2026)✨

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#usage使用

Cohere Transcribe Arabic 在 transformers 中原生支持。这是推荐用于离线推理的使用方式。在线推理请参见下面的 vLLM 集成示例。

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

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#quick-start-%F0%9F%A4%97快速入门 🤗

几行代码即可转录任意音频文件：

`` 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”)

示例：转录阿拉伯语音频

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) ``

长音频转录对于超过特征提取器 max_audio_clip_s 的音频，特征提取器会自动将波形分割成片段。处理器使用返回的 audio_chunk_index 重新组装每个片段的转录。

`` 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“音频时长: {duration_s / 60:.1f} 分钟“)

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“转录耗时 {elapsed:.1f}s — RTFx: {rtfx:.1f}“) print(text) ``

英语转录该模型也支持英语。指定 language="en" 即可：

`` 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) ``

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#vllm-integrationvLLM 集成

对于生产环境服务，我们建议按照以下说明通过 vLLM 运行。

通过 vLLM 运行 cohere-transcribe-arabic-07-2026首先安装 vLLM（参考 vLLM 安装说明 (https://docs.vllm.ai/en/latest/getting_started/installation/)）：

`` 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 ``

启动 vLLM 服务器

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

发送请求

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"

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#results结果

开放通用阿拉伯语 ASR 排行榜（截至 2026.07.07）模型平均 WER · CERSADWER · CERCommon VoiceWER · CERMASC 干净WER · CERMASC 嘈杂WER · CERMGB-2WER · CERCasablancaWER · CERCohere Transcribe Arabic 07-202625.8711.8037.4723.535.821.6219.606.4527.0710.1315.548.4049.71****20.66OmniASR LLM 7B28.3212.5241.6124.958.752.7119.695.7629.2910.6614.137.1056.4623.96OmniASR LLM 3B29.9613.7746.1827.279.152.8019.906.1330.0311.2714.227.0660.2728.06OmniASR LLM 1B29.9613.4043.8424.549.552.9720.036.1430.2611.1815.347.5660.6828.02Cohere Transcribe 03-202630.6716.3760.1145.448.172.498.662.9719.01****7.7125.339.2862.7130.31Qwen3-Omni 30B30.7113.6744.8226.1111.464.2821.475.5930.8511.2813.09****6.2062.5528.53NVIDIA Conformer-CTC (LM)32.9113.8444.5223.768.802.7723.745.6334.2911.0717.206.8768.9032.97OmniASR LLM 300M32.9614.8451.3829.1012.034.0420.666.2232.4512.2316.587.8664.6429.61Gemma 4 E4B32.9813.7143.4020.9619.657.4824.867.7633.5912.2517.728.6758.6325.11Qwen3-ASR 1.7B33.3612.3345.5319.9016.905.0624.375.7234.2910.8416.576.2564.4726.23Voxtral-Small 24B34.4715.2950.8228.8515.255.5423.967.0634.4312.2216.037.4166.3030.64NVIDIA Conformer-CTC (贪婪)34.7413.3747.2622.5410.603.0524.125.6335.6411.0219.697.4671.1330.50Gemma 4 E2B35.8715.3446.2323.4723.769.1327.478.9936.1513.9320.7210.1560.8726.35Whisper Large v336.8617.2155.9634.6217.835.7424.667.2434.6312.8916.267.7471.8135.04 实时排行榜链接：开放通用阿拉伯语 ASR 排行榜 (https://huggingface.co/spaces/elmresearchcenter/open_universal_arabic_asr_leaderboard)。

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#resources资源

更多详情和结果：

• 技术博客文章 (https://huggingface.co/blog/CohereLabs/cohere-transcribe-arabic-07-2026-release)包含 WER 及其他质量指标。
• 公告博客文章 (https://cohere.com/blog/transcribe-arabic)提供有关该模型的更多信息。
• 开放通用阿拉伯语 ASR 排行榜 (https://huggingface.co/spaces/elmresearchcenter/open_universal_arabic_asr_leaderboard)。

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#strengths-and-limitations优势与局限

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#strengths优势

Cohere Transcribe Arabic 在阿拉伯语和英语的转录准确性上表现出色。作为专用的语音识别模型，它凭借 Conformer 编码器-解码器架构实现了高效的推理。

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#limitations局限

• 单一语言：该模型在保持单一预设语言分布内时表现最佳。它不具备显式的自动语言检测功能，且在代码切换音频上表现不一致。
• 时间戳/说话人分离：该模型不具备这两项功能。
• 静音：与大多数 AED 语音模型类似，Cohere Transcribe Arabic 对于转录（甚至是非语音声音）过于积极。建议在模型前加入噪声门或 VAD（语音活动检测）模型，以防止低音量底噪导致幻听。

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#model-card-contact模型卡联系方式

关于此模型卡中的错误或更多问题，请联系 [email protected] 或提交 issue。

使用条款：我们希望此模型的发布能使社区对阿拉伯语语音的研究工作更加便捷。该模型采用 Apache 2.0 许可证。

https://huggingface.co/CohereLabs/cohere-transcribe-arabic-07-2026#citation引用

引用此模型请使用以下 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 } }