Salesforce’s SOV-Transcribe-1 Ranks First Among Open-Weight Speech Models

Salesforce’s SOV-Transcribe-1 scored 2.3% on a speech-recognition benchmark, placing seventh among 44 systems overall and first among models whose weights users can download for local deployment.

The ranking included both proprietary systems and open-weight models. Mistral AI’s Voxtral Small and NVIDIA’s Nemotron 3.5 Transcribe each scored 2.8%, 0.5 percentage point above Salesforce’s model.

On three separate test sets, SOV-Transcribe-1 recorded a 4.1% word error rate on Earnings22, 1.5% on VoxPopuli and 1.8% on AA-AgentTalk.

The model has 26.6 billion parameters, with 4.6 billion active at a time. Salesforce published the model weights on Hugging Face for organizations to deploy within infrastructure they control. That deployment option may suit organizations with strict data-control requirements.

Salesforce placed the release under its Sovereign AI initiative and tied it to Missionforce, its effort focused on government and national-security workloads. Salesforce launched Missionforce in September 2025.