Home / Technology / Open-Weight AI Beats Whisper in Accuracy
Open-Weight AI Beats Whisper in Accuracy
30 Mar
Summary
- Cohere's Transcribe achieves 5.42% word error rate.
- Transcribe offers local deployment for data control.
- Model outperforms Whisper and ElevenLabs on accuracy.
Cohere has launched Transcribe, an open-weight automatic speech recognition (ASR) model designed to offer production-grade transcription for voice-enabled workflows. This model aims to compete on key differentiators including contextual accuracy, latency, control, and cost.
Transcribe boasts a 2 billion parameter size and is licensed under Apache 2.0. The model has demonstrated an average word error rate (WER) of 5.42%, outperforming industry benchmarks such as OpenAI's Whisper Large v3 (7.44%) and ElevenLabs Scribe v2 (5.83%). It supports 14 languages, including English, French, German, Italian, Spanish, and others.
Previously, enterprises faced a trade-off between accurate but data-risky closed APIs and less accurate open models. Transcribe offers a solution by providing high accuracy and the ability to run on an organization's local GPU infrastructure, ensuring data control and addressing residency concerns. This commercial-ready approach marks a significant advancement for enterprise deployments.
The model's manageable inference footprint makes local deployment feasible. Cohere states Transcribe extends the Pareto frontier by delivering state-of-the-art accuracy with high throughput. Early users have highlighted the model's accuracy and local deployment capabilities as key advantages for bringing audio workloads in-house.
