Salesforce has released the model weights for SOV-Transcribe-1: a 26.6-billion-parameter speech transcription model that activates 4.6 billion parameters per inference, supports audio up to 30 minutes, and ranks first among open-weight models on Artificial Analysis’s non-streaming speech recognition leaderboard. We believe the key shift is that, for enterprises deploying speech AI, the priority is moving from “which cloud API to call” toward “whether we can keep the data and model in our own hands.”
What this is
SOV-Transcribe-1 is an automatic speech recognition model (ASR, which converts speech signals into text). Beyond English, it supports six languages, including French, German and Spanish. It can handle accents and background noise, while natural-language instructions let users choose between verbatim transcripts and edited text, with control over punctuation, capitalization and number formatting. It can also be supplied with company-specific terms to improve recognition of product names and people’s names.
The model is available under the Apache 2.0 license and can be deployed with vLLM and Hugging Face Transformers. Salesforce also provides the DFlash speculative decoding model to accelerate inference by generating candidate text first. Open weights do not mean that it works out of the box: enterprises still need to prepare computing resources, inference services and ongoing engineering support.
Industry view
Our assessment is that competition in speech models is expanding beyond accuracy alone to include control and deployment options. Natural-language control over transcription style and adaptation to proprietary terminology offer practical value for customer service, medical records and sales follow-up, among other use cases. Keeping audio off third-party platforms also directly addresses enterprise concerns about data sovereignty and privacy.
The risks are equally clear. Leaderboard results come from a specific test set and testing period and cannot be assumed to apply to every enterprise recording environment. The model’s language coverage excludes Chinese, even though Chinese-language customer service, meetings and dictation are core requirements for many businesses. Although the model can be deployed privately, enterprises remain responsible for hardware costs, model evaluation, monitoring and upgrades. Moreover, official claims about an open model do not mean Salesforce has already demonstrated large-scale commercial deployment.
Impact on regular people
For enterprise IT: Organizations with meeting transcripts, customer-service quality assurance or internal speech archives can evaluate local deployment, testing their actual requirements around proprietary terminology, noise and Chinese-language coverage before deciding whether to purchase servers or cloud computing capacity.
For individual professionals: Speech transcription will become a more reliable part of meeting, interview and customer-service workflows, but the lack of Chinese-language support means the direct benefit for many individuals may be limited. Higher-quality results will still depend on recording equipment and domain-specific terminology.
For the consumer market: Open weights could lead meeting software, recording apps and captioning tools to add private deployment options. In the short term, however, the technology is more likely to appear first in enterprise services rather than become a free feature available to everyone.