Salesforce has released SOV-Transcribe-1, a new speech-to-text model currently ranked as the top open-weights option on the Artificial Analysis Word Error Rate (AA-WER) Index. The model achieved a 2.3% word error rate, placing it seventh overall among 44 evaluated systems, including proprietary ones.
The model features 26.6 billion total parameters, with 4.6 billion active at any given time. According to research findings, the model’s host endpoint speed factor is approximately 60 times faster than conventional systems. It also recorded a 4.1% word error rate on the Earnings22 dataset and 1.5% on VoxPopuli.
Available on Hugging Face since November 12, 2026, the model allows organizations to run it on their own infrastructure. It outperforms rivals such as Mistral AI’s Voxtral Small and NVIDIA’s Nemotron 3.5 Transcribe, which both posted a 2.8% word error rate.
SOV-Transcribe-1 is part of Salesforce’s Sovereign AI initiative and integrates with Missionforce, a project launched in September 2025. This initiative focuses on providing secure, controllable AI for government and enterprise clients who require local deployment for data compliance.
