
SOV-Transcribe-1: Salesforce's Open Speech Model Explained
Salesforce's SOV-Transcribe-1 is an Apache 2.0 open-weight speech-to-text model with a 2.28% word error rate, style steering and 7 languages.
Speech recognition has quietly become one of the most competitive corners of open AI, and a new entrant just landed near the top. On November 12, Salesforce's Sovereign Intelligence team released SOV-Transcribe-1, an open-weight speech-to-text model under the Apache 2.0 license. Salesforce says it is built for the messy audio real businesses deal with: call centers, crowded subways, heavy accents and dictated email addresses.
- Size: 26.6B total parameters with about 4.6B active, pairing a Whisper large-v3 audio encoder with a Gemma 4 26B A4B mixture-of-experts decoder
- Accuracy: a 2.28% word error rate on Artificial Analysis's benchmark, per Salesforce, ranking first among open-weight non-streaming models and seventh overall
- Languages: English, French, German, Spanish, Italian, Portuguese and Dutch, with audio inputs tested up to 30 minutes
- License: Apache 2.0, with Hugging Face Transformers and vLLM support
What Makes SOV-Transcribe-1 Different?
Most automatic speech recognition models simply write down what they hear. SOV-Transcribe-1 is a speech-augmented language model, which means it accepts text instructions alongside the audio. That unlocks two features developers will notice right away.
Style steering lets you decide how the transcript looks: verbatim with every "um," or clean written-form text with formatting. Context biasing lets you feed in names, domain vocabulary or a short topic description so the model gets difficult terms right, the kind of thing that matters when a transcript feeds a medical record or a sales system.
How Accurate Is Salesforce's Open Speech Model?
All benchmark figures here come from Salesforce's own launch post, so treat the comparisons as the company's claims. On Artificial Analysis's word error rate test, Salesforce reports 2.28%, ahead of NVIDIA's Nemotron 3.5 Transcribe at 2.81% with a similar active size and Mistral's Voxtral Small at 2.77%. Other open models of similar size sit around 4%, according to the post.
The more telling numbers come from entity accuracy. In a voice agent, one swapped digit in an account number can break an entire task. On VoiceCodeBench, which only passes a recording if every target entity is correct, Salesforce says the model passes 76.3% of recordings and recovers 93.4% of entities, ranking second of 23 systems. On a long-form call-center test spanning 14 English accents, it posts an 8.21% average error rate and the lowest score on eight of those accents, including Scottish, Irish, Indian and Singaporean English.
Can You Run SOV-Transcribe-1 Yourself?
Yes, and that is the point of an open-weight release. The model card includes setup for Transformers and a vLLM plugin with an OpenAI-compatible transcription endpoint. Salesforce says a single RTX PRO 6000 Blackwell GPU transcribes up to 1,039 seconds of audio per second and returns a 30-second clip in 0.22 seconds. This is server-class hardware rather than a laptop model, but enterprises that want voice AI without shipping audio to a third party now have a strong, permissively licensed option.
Why Open-Weight Speech-to-Text Keeps Getting Better
SOV-Transcribe-1 arrives just weeks after NVIDIA released Nemotron 3.5 Transcribe, and both lean on large language model decoders to understand context instead of just matching sounds. The result is a fast-moving race where each release raises the floor for everyone building transcription, captioning and voice agents. Follow it with the rest of our AI coverage.
Sources: Salesforce Sovereign Intelligence on Hugging Face — November 12, 2026; SOV-Transcribe-1 model card — November 2026; Artificial Analysis speech-to-text leaderboard — accessed November 15, 2026.
More AI Stories

Gemini Google Wallet Integration: AI Budgeting Explained
Gemini can now read your Google Wallet spending to chart budgets, flag pricier subscriptions and plan holiday costs. Here's how the US rollout works.

XYBench: Why AI Assistants Miss the Real Question You Ask
Ai2's XYBench tests 8,115 misguided questions to see if AI assistants spot the real problem. The award-winning benchmark shows where models can improve.

Google AI Edge Gallery Adds Offline Speech: What Devs Get
Google AI Edge Gallery now runs speech recognition and text-to-speech fully on device, with 100+ languages and replies that start in about 100 ms.
