Salesforce releases SOV-Transcribe-1, now the top open-weights speech model

The model posted a 2.3% word error rate on the AA-WER Index, placing it seventh out of 44 evaluated models. This performance puts the CRM company ahead of several AI labs whose entire identity is building models. Salesforce unveiled SOV-Transcribe-1 on November 12, 2026, making the weights available on Hugging Face for organizations to download and run on their own infrastructure.

Performance metrics

The closest open-weights rivals are Mistral AI's Voxtral Small and NVIDIA's Nemotron 3.5 Transcribe, both scoring 2.8% on the AA-WER Index. SOV-Transcribe-1 also posted leading results on specific test sets, scoring 4.1% WER on Earnings22 and 1.5% WER on VoxPopuli. On AA-AgentTalk, another evaluation component, the model achieved 1.8%.

Technical specifications

Under the hood, SOV-Transcribe-1 carries 26.6 billion total parameters, with only 4.6 billion active at any given time. The research findings indicate the model's host endpoint speed factor is approximately 60 times that of conventional systems, making it highly efficient despite its size.

Sovereign AI initiative

The model is part of Salesforce's Sovereign AI initiative, connecting to Missionforce, the company's effort to deliver secure, controllable AI for enterprise and government customers. Missionforce launched in September 2025, focusing on government and national security workloads where sensitive audio cannot be sent to external servers.

For organizations with strict data control and compliance rules, an open-weights model that can be deployed internally is not just a preference but a requirement, addressing critical security concerns in sensitive sectors.