Cohere Transcribe Claims Top Spot on ASR Leaderboard
Cohere has released its new automatic speech recognition (ASR) model, Cohere Transcribe, which has achieved the number one ranking on the Hugging Face Open ASR Leaderboard. Launched on March 26, 2026, the model achieved an average Word Error Rate (WER) of 5.42%. This positions it ahead of established models such as Whisper Large v3, which has a WER of 7.44%, and IBM Granite 4.0 at 5.52%.
Cohere Transcribe was trained using standard supervised cross-entropy. The model supports 14 languages and performs best when the target language is pre-defined, lacking explicit automatic language detection or optimized support for code-switching. It is a pure ASR tool without native speaker diarization or timestamps.
Innovative Architecture Enhances Performance
The design of Cohere Transcribe features a hybrid architecture, employing a large Conformer encoder in conjunction with a lightweight Transformer decoder. This combination is engineered to effectively process both local acoustic features through convolution and global linguistic context via self-attention. This differs from standard pure-Transformer models by integrating elements typically associated with Convolutional Neural Networks (CNNs).
To address the challenges of memory-intensive architectures when handling long-form audio, Cohere Transcribe utilizes a native 35-second chunking logic. For audio files exceeding this duration, the system segments them into overlapping chunks. Each segment is processed, and the overlapping text is then reassembled to maintain continuity.
Long-Form Audio and Human Preference Metrics
Cohere Transcribe demonstrates robust performance with extended audio files. The model can process a 55-minute file without exhausting GPU VRAM, a significant advantage for tasks involving extensive recordings. This is achieved through its chunking and reassembly logic, avoiding typical sliding-window attention methods.
In comparative evaluations, annotators showed a strong preference for transcripts generated by Cohere Transcribe. Head-to-head comparisons revealed that annotators preferred Transcribe over competing transcripts by 78% against IBM Granite 4.0 1B Speech, 67% against NVIDIA Canary Qwen 2.5B, 64% against Whisper Large v3, and 56% against Zoom Scribe v1.
✨ Intelligent Curation Note
This article was processed by AI Universe’s Intelligent Curation system. We’ve decoded complex technical jargon and distilled dense data into this high-impact briefing.
Estimated time saved: ~2 minutes of reading.
Tools We Use for Working with AI:
