Open speech recognition has evolved beyond the Whisper monoculture in 2026. In March 2026, Cohere released Transcribe, a 2-billion-parameter model under the Apache 2.0 license. It quickly topped the Hugging Face Open ASR Leaderboard with an average Word Error Rate (WER) of just 5.42%. Just five weeks later, a new contender emerged...
Key Metrics at a Glance (2026)
| Model | WER (%) | Languages | Latency (ms) | License |
|-------|---------|-----------|--------------|---------|
| Cohere Transcribe | 5.42 | 100+ | 120 | Apache 2.0 |
| Whisper Large-v3 | 6.10 | 99 | 210 | MIT |
| Nvidia NeMo Canary | 5.89 | 60+ | 95 | Apache 2.0 |
| Meta SeamlessM4T-v2 | 6.35 | 100+ | 180 | CC BY-NC 4.0 |
Word Error Rate (WER) Comparison
WER remains the primary benchmark for ASR accuracy. As of mid-2026, Cohere Transcribe leads with a 5.42% WER, outperforming both Whisper Large-v3 (6.10%) and Meta's SeamlessM4T-v2 (6.35%). Nvidia's NeMo Canary offers competitive accuracy (5.89%) with lower latency.
Language Support
All leading models support over 60 languages. Cohere Transcribe and SeamlessM4T-v2 both exceed 100 languages. Whisper supports 99 languages, while NeMo Canary covers 65 languages. For enterprise use, the practical difference narrows: all models handle major languages well, but coverage for lower-resource languages varies significantly.
Latency and Deployment
Latency is critical for real-time applications. NeMo Canary leads with 95ms average latency, followed by Transcribe at 120ms. Whisper Large-v3 (210ms) and SeamlessM4T-v2 (180ms) are slower but still suitable for asynchronous tasks. Optimized inference engines (e.g., TensorRT, ONNX) can reduce latency by 30-50% across models.
Licensing Considerations
Cohere Transcribe (Apache 2.0) and NeMo Canary (Apache 2.0) are fully open for commercial use. Whisper Large-v3 uses the MIT license, also liberal. Meta's SeamlessM4T-v2 uses CC BY-NC 4.0, which prohibits commercial use unless separately licensed. Teams should verify licensing terms for production deployments.
Conclusion
The 2026 open ASR landscape offers multiple strong options. For lowest WER, choose Cohere Transcribe. For fastest real-time inference, Nvidia NeMo Canary is best. For broadest language support and research flexibility, Whisper Large-v3 remains a reliable baseline. Always check licensing for your specific use case.
Data sourced from Hugging Face Open ASR Leaderboard and model release documentation, accessed July 2026.
via MarkTechPost