Salesforce releases a voice to text model with a 2.3% word error rate
With open weights, a team can get the model and operate it on its computers. This is important for audio that no one else may hear.
Salesforce, which makes customer software, released SOV-Transcribe-1, a voice to text model, on 12 November 2026. On the Artificial Analysis word error rate index, it has a score of 2.3%. This is first of the models with open weights and number 7 of 44 models in all. The weights are on Hugging Face. A team can operate the model on its computers.
How it works
The model has 26.6 billion parameters, the numbers in the model. Only 4.6 billion are active at one time. Salesforce puts it in Sovereign AI and in Missionforce, which started in September 2025 for government and national security customers.
The numbers
The next 2 models with open weights, from Mistral AI and NVIDIA, each have a score of 2.8%. The research results put the speed of the host endpoint at about 60 times the speed of usual systems.
What is not known
The price and the license are not given. Salesforce gave no other system to compare with the 60 times number. The index is one test, and other tests can put the models in a different sequence.
Insight: Salesforce now builds its small models
Salesforce also told of Koa Decisions, a routing model made from the open Gemma model of Google. At this time it releases a voice model with open weights in Sovereign AI. This reads as a plan to hold more of the model layer below its agent software, and to give control to government customers. Look for more open models from Salesforce for Missionforce customers.
Sources
Posted
