THE ESSENTIALS IN 30 SECONDS
- SOV-Transcribe-1, released November 12, 2026, transcribes English, French, German, Spanish, Italian, Portuguese and Dutch.
- The mixture-of-experts model has 26.6 billion parameters, 4.6 billion of them active per token, and ships under Apache 2.0.
- It reports a 2.28% word error rate on Artificial Analysis's index, first among open-weight models; most other figures come from Salesforce.
Salesforce has entered the open speech-model field with SOV-Transcribe-1, an automatic speech recognition model whose weights were posted on Hugging Face on November 12, 2026. It comes from the company's Sovereign Intelligence team and is released under the permissive Apache 2.0 license. The "sovereign" label refers to the main selling point: organizations can download the model and run it on servers they control, rather than sending audio to an external transcription service.
What the model is
SOV-Transcribe-1 combines an audio encoder derived from OpenAI's Whisper large-v3 with a text decoder based on Google's Gemma 4 26B A4B, a mixture-of-experts model. The result has 26.6 billion parameters in total, of which about 4.6 billion are active for each token. It handles seven languages: English, French, German, Spanish, Italian, Portuguese and Dutch.
The model works on recorded files of up to 30 minutes, sampled at 16 kHz; it does not transcribe live streams. Users can give plain-language instructions to choose between a verbatim transcript, hesitations included, and cleaner prose. They can also supply glossaries or a short topic summary so that names and specialist terms come out right without retraining. It can be served with Hugging Face Transformers or with vLLM through a plugin, and a smaller draft model is offered to speed up decoding.
Benchmark claims
On the word error rate index of the independent firm Artificial Analysis, SOV-Transcribe-1 scores 2.28%. That puts it first among open-weight models and seventh overall when proprietary services are included, according to TokenPost and CASRAI. The same comparisons place Nvidia's Nemotron 3.5 Transcribe and Mistral's Voxtral Small at around 2.8%. Salesforce also cites results on a call-center test covering 14 English accents and on an entity-recognition benchmark; those figures are the company's own and have not been independently checked.
Context and open questions
Salesforce links the release to its Missionforce effort aimed at government and national-security customers, a unit it set up in September 2025. Even with only a fraction of its parameters active at once, all 26.6 billion must be loaded, so running it locally still calls for substantial GPU memory. Real-world accuracy on noisy, domain-specific audio will only become clear as outside users test it.
