Salesforce Sovereign Intelligence has published SOV-Transcribe-1, an Apache-2.0 open-weight speech-to-text model aimed at non-streaming transcription where mistakes in a name, number, email address or technical term can invalidate an entire workflow. The November 12 release includes a downloadable Hugging Face checkpoint and examples for developer-controlled inference. The model is designed not just to convert audio into words but to follow transcription instructions, such as preserving fillers for a verbatim record or formatting numbers as digits. Those capabilities make it relevant to enterprise applications, although on-premises execution alone is not proof of privacy compliance. Primary source ↗
The official model card gives a more precise size than the rounded launch summary: 26.6 billion parameters in total, 4.6 billion active, including a Whisper large-v3 audio encoder, projector and Gemma 4 26B-A4B mixture-of-experts language backbone. It supports 16 kHz input and seven specified languages: English, French, German, Spanish, Italian, Portuguese and Dutch. The team reports testing audio clips up to 30 minutes. SOV-Transcribe-1 is non-streaming; it should not be described as a real-time duplex voice assistant or a tiny on-device mobile model.
Salesforce emphasizes entity-level accuracy and domain-context biasing. Its announcement reports 2.28% word error rate in an Artificial Analysis comparison and 76.3% task success with 93.4% entity recovery on VoiceCodeBench. The published rankings are tied to named benchmark configurations and the November 12 comparison set. The team also reports a 0.22-second transcript for a 30-second recording with a Blackwell-class server GPU and speculative draft model. These are publisher-reported figures; Open Model Weights has not independently verified the scoring, acceleration path or throughput under customer audio.
Deployment caveats matter as much as the headline. The official Transformers example uses trust_remote_code and the vLLM instructions rely on a bundled plugin pending upstream support. Security teams should review executed modeling code, pin revisions, check the separate DFlash draft repository and measure memory and latency in their own setup. The registry should preserve the exact Apache license declaration and distinguish supported languages, measured audio-length limit, benchmark methodology and runtime plugin requirements. For speech agents, a robust transcript is an input to action, not authorization to execute that action.
What this changes for the evidence layer
Check salesforce-sovereign-intelligence/SOV-Transcribe-1 canonical SHA, Apache-2.0 and included code licensing, 26.6B/4.6B parameters, seven declared languages, 30-minute tested limit, custom code / vLLM plugin and DFlash separate checkpoint. Attribute benchmark and throughput measurements to Salesforce; no independent performance ranking.
Salesforce Sovereign Intelligence — SOV-Transcribe-1 official model card
This brief is based on the cited primary source. Performance, benchmark and comparative claims remain attributed unless Open Model Weights publishes an independent measurement.
Open source article ↗