The new transcription model posts a 2.3% word error rate and ranks seventh overall among 44 models on the AA-WER Index
Salesforce has a new speech-to-text model, and it is currently the best open-weights option on the board. SOV-Transcribe-1 ranks first among open-weights speech-to-text models and seventh overall on the Artificial Analysis Word Error Rate (AA-WER) Index. That puts a CRM company ahead of several AI labs whose entire identity is building models.
The model posted a 2.3% word error rate on the index. That score placed it seventh out of 44 evaluated models, a field that includes proprietary systems from companies that do not hand out their weights.
What Salesforce actually shipped #
Salesforce unveiled SOV-Transcribe-1 on November 12, 2026. The weights are available on Hugging Face, so organizations can download the model and run it on their own infrastructure.
The closest open-weights rivals are Mistral AI’s Voxtral Small and NVIDIA’s Nemotron 3.5 Transcribe. Both scored 2.8% on the AA-WER Index.
SOV-Transcribe-1 also posted leading results on two specific test sets within the evaluation. It scored a 4.1% WER on Earnings22 and a 1.5% WER on VoxPopuli.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
On AA-AgentTalk, another component of the evaluation, the model hit 1.8%.
Big model, smaller footprint in practice #
Under the hood, SOV-Transcribe-1 carries 26.6 billion total parameters. Only 4.6 billion of those are active at any given time.
The research findings put the model’s host endpoint speed factor at approximately 60 times that of conventional systems.
The sovereign AI angle #
The model sits under Salesforce’s Sovereign AI initiative. It ties into Missionforce, the company’s effort to deliver secure, controllable AI for enterprise and government customers.
Missionforce launched approximately one year earlier, in September 2025. Its focus is government and national security workloads, a market where nobody gets excited about sending sensitive audio to someone else’s server.
An open-weights model can be deployed inside an agency’s own environment. For organizations with strict data control and compliance rules, that is often not a preference but a requirement.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.