Salesforce SOV-Transcribe-1: open-weight speech-to-text, Apache 2.0

Salesforce now gives away a speech model you can run on your own hardware. SOV-Transcribe-1 is an open-weight speech-to-text model with an Apache 2.0 license, released on Hugging Face on November 12. Salesforce claims the top spot among open-weight models for batch transcription. The numbers are the vendor's own. The license is the part you can bank on.

What actually happened

• The model. The model card lists 26.6 billion parameters in total and 4.6 billion active: a Whisper large-v3 audio encoder on a Gemma 4 decoder. It covers seven languages and recordings up to 30 minutes. It does non-streaming speech recognition only. You send a recording and get a transcript. It doesn't do live captions.
• The headline number. Salesforce cites a 2.28% word error rate on the Artificial Analysis benchmark, against 2.81% for NVIDIA's Nemotron 3.5 Transcribe and 2.77% for Mistral's Voxtral Small. It ranks first among open-weight models there and seventh overall.
• The number that matters more. On VoiceCodeBench, which checks phone numbers, emails, URLs and code symbols, the model gets every entity right in 76.3% of recordings. That means about one recording in four still has a wrong digit or a dropped character.
• The hard audio. On call-center dialogues across 14 English accents, the average error rate is 8.21%. On noisy earnings calls picked to cause hallucinations, it's 25.69%. That's the best of the open-weight group and still one word in four.
• The controls. You steer the format with plain instructions, such as "write numbers as digits." You can also pass keywords. With a clip's keywords supplied, name accuracy (keyword F1) went from 71.9 to 96.8 in Salesforce's test.

Salesforce says a technical report is "coming soon." Until then, these are claims from the party that sells the product.

Why it matters for your business

If you record sales calls, support calls or job-site voice notes, you pay per audio hour to a cloud API and you send customer audio off your network. An Apache 2.0 model removes both. You can run it, change it and ship it in a product with no usage fee.

The cost moves to hardware. A model with 26.6 billion parameters needs a serious GPU, so price that before you cancel an API plan.

Our advice:

1. Test on your own audio. Take 50 real calls. Score names, phone numbers and order numbers, not the average word error. The 2.28% figure came from clean benchmark audio.
2. Feed it your vocabulary. The keyword result is the best number in the release. Pass your product names, customer names and SKUs with each job.
3. Keep it batch. Use it for recorded calls and voicemail. For a live voice agent, you need a streaming model.
4. Keep the output portable. Store transcripts in your own database with the audio file ID. Then you can swap models when the next one ships. One will.

Sources: Salesforce on Hugging Face, SOV-Transcribe-1 model card.