Mini toolsCohere Transcribe Arabic
By Cohere
Cohere Transcribe Arabic is a 2B parameter dedicated automatic speech recognition model built on a conformer based encoder decoder architecture. A large Conformer encoder extracts acoustic representations from a log Mel spectrogram, which a lightweight Transformer decoder converts into transcribed text. The model accepts audio waveforms resampled to 16kHz and outputs text in either Arabic or English, with support for long form audio through automatic chunking. It was trained with a supervised cross entropy objective on output tokens and is released under the Apache 2.0 license with native support in the transformers library and vLLM for production serving. On the Open Universal Arabic ASR Leaderboard the model achieves the lowest average word error rate among tested systems across benchmarks such as SADA, Common Voice, MASC, MGB2, and Casablanca. It performs best within a single specified language and does not support timestamps, speaker diarization, or non speech sound suppression.
New
Audio
Overview
Cohere Transcribe Arabic is a 2B parameter conformer based encoder decoder automatic speech recognition model that transcribes Arabic and English audio into text. It achieves strong accuracy across Arabic dialect benchmarks including SADA, Common Voice, MASC, MGB2, and Casablanca, outperforming several larger open ASR models.
About Cohere
Cohere is a technology company specializing in Artificial Intelligence, Natural Language Processing, and Machine Learning solutions.
Industry:
Artificial Intelligence
Company Size:
900
Location:
Toronto, Ontario, CA
Website:
cohere.com
Last updated: July 8, 2026
