Salesforce has introduced SOV-Transcribe-1, an open-weight speech recognition model designed for non-streaming transcription. In its November 12 announcement, the team describes a model with 26 billion total parameters and four billion active parameters, alongside controls for transcription style and terminology.
More than a transcript
The publisher says users can request verbatim or cleaned-up text and provide context such as names or specialized vocabulary. Those features are useful in theory because a single incorrect name, address or number can change the meaning of a transcript. They do not remove the need to check important output against the original audio.
The model card identifies Apache 2.0 licensing for the weights and describes the supported tools. Salesforce reports results on transcription and entity benchmarks, including comparisons with other models; those are publisher-presented findings with dataset and measurement limits. We have not reproduced the results.
Local does not necessarily mean small
This is a mixture-of-experts model. Four billion active parameters per inference step do not mean only four billion parameters need storage or memory. Review the model card’s actual hardware requirements and software instructions before planning a local deployment, especially on consumer GPUs.
Because the release is non-streaming, it also should not be assumed to deliver live word-by-word captions. Organizations handling sensitive recordings should review the complete deployment, dependencies and data-handling practices rather than assuming that open weights alone guarantee privacy.
Original source-linked reporting; no independent product or benchmark test.
