Salesforce drops SOV-Transcribe-1, the best open-weights speech model yet

Salesforce has a new speech-to-text model, and it is the best open-weights option on the board. SOV-Transcribe-1 ranks first among open-weights models and seventh overall on the Artificial Analysis Word Error Rate Index, out of 44 evaluated systems. That overall field includes proprietary models from companies that keep their weights locked up.

The model posted a 2.3% word error rate on the index. Its closest open-weights competitors are Mistral AI's Voxtral Small and NVIDIA's Nemotron 3.5 Transcribe, both at 2.8%. Salesforce unveiled SOV-Transcribe-1 on November 12, 2026, and published the weights on Hugging Face so organizations can download and run it on their own infrastructure.

On specific test sets, SOV-Transcribe-1 scored a 4.1% WER on Earnings22 and 1.5% on VoxPopuli. It also hit 1.8% on AA-AgentTalk, another component of the evaluation. Under the hood it carries 26.6 billion total parameters, with only 4.6 billion active at any given time. Salesforce's research puts its host endpoint speed factor at roughly 60 times that of conventional systems.

The model falls under Salesforce's Sovereign AI initiative and ties into Missionforce, the company's push to deliver secure, controllable AI for enterprise and government customers. Missionforce launched about a year earlier, in September 2025, aimed at government and national security workloads. For agencies with strict data control and compliance rules, running a model inside their own environment is less a preference than a requirement.