Skip to content
ReviewRamble

The weekly brief

AI, reviewed, in your inbox.

  • New reviews of tools, models, MCP servers and repos, with the verdict up front
  • Re-checked prices and ratings when something changes
  • New analysis pieces as they are published

Models, agents, MCP servers and the tools actually worth your time.

One email a week. Unsubscribe with the link in any brief. How we use your email

Salesforce SOV-Transcribe-1

Salesforce's Apache 2.0 transcription model from November 2026: the most accurate open weights on Artificial Analysis, at 2.3% error.

Reviewed by ReviewRamble editors · Updated

Share this page:XLinkedInFacebookRedditWhatsAppEmail

Our rating

4.3/5

Open transcription that finally sits among the commercial leaders, though it wants a 96 GB GPU and speaks only seven languages.

What is Salesforce SOV-Transcribe-1?

SOV-Transcribe-1 is an open-weight speech recognition model released by Salesforce on 12 November 2026 under the Apache 2.0 licence. It joins a Whisper large-v3 audio encoder to a Gemma 4 26B A4B mixture-of-experts language model: 26.6 billion parameters in total, about 4.6 billion active at each step. It transcribes recorded audio, not live streams.

Because the decoder is a language model, you can give it formatting instructions (verbatim for a deposition, clean text for a dictated message) and a list of key terms such as customer or drug names. The model card lists English, French, German, Spanish, Italian, Portuguese and Dutch, and recordings of up to 30 minutes as tested.

It runs on Hugging Face Transformers or vLLM through a plugin shipped in the repository, with an optional draft model for speculative decoding. Salesforce frames it as a building block for companies that want voice features on their own infrastructure.

Is Salesforce SOV-Transcribe-1 worth it?

SOV-Transcribe-1 closes most of the gap between what you can download and what you can rent. Artificial Analysis measures it at 2.3% word error rate, the best of any open-weight model and within 0.4 points of ElevenLabs' paid Scribe v3. Salesforce's own figure is 2.28%, against 2.81% for Nvidia's Nemotron 3.5 Transcribe and 2.77% for Mistral's six-times-larger Voxtral Small.

The more useful numbers are about the words that matter. On VoiceCodeBench, which checks emails, phone numbers and similar entities, it ranked second of 23 systems and recovered 93.4% of entities. On noisy earnings calls, where speech models tend to invent text, it scored 25.69% against Nemotron's 31.59%. Those comparisons are Salesforce's, and the technical report is still marked as coming soon.

The catch: this is a large model for one job. All of Salesforce's benchmarks ran on a single 96 GB RTX PRO 6000 Blackwell, and the default vLLM settings run out of memory there. Support for Transformers and vLLM is not yet upstream, so you depend on a plugin pinned to vLLM 0.32.0. It cannot stream, and the card warns that it occasionally loops on a phrase until the token limit, so outputs need a repetition filter.

Our take: for a company that must keep call audio in-house in English or a major Western European language, this is now the open model to beat. A team that needs live captions, Asian languages or a laptop-sized footprint should look at Nemotron or Parakeet instead.

Where it earns its place

  • A bank's compliance team recording advisory calls

    Transcribe every call on premises with client names and account numbers right

    Serve the model with vLLM on one 96 GB GPU and pass each adviser's client list as key terms. Salesforce reports peak throughput of 1,039 seconds of audio per second.

  • A court-reporting firm in the Netherlands and Belgium

    Produce verbatim Dutch and French transcripts of depositions

    Ask for verbatim output with greedy decoding, split recordings over 30 minutes at pauses and run a repetition check before review.

Before you commit

Setup
pip install transformers (5.19 to 5.22), download the vllm_plugin folder, pip install it, then vllm serve salesforce-sovereign-intelligence/SOV-Transcribe-1 --max-model-len 32768; send 16 kHz mono audio with temperature 0.
Integrations
Hugging Face Transformers (trust_remote_code) and an OpenAI-style /v1/audio/transcriptions endpoint through vLLM; optional DFlash draft model for speculative decoding.
Data and privacy
Self-hosted under Apache 2.0, so audio stays on your servers. Training data and full training code are not part of the release.
Cost at scale
No per-minute fee. With the draft model a 30-second clip takes about 0.22 s on one RTX PRO 6000 Blackwell; cost is that GPU's time.
Limits
No streaming. vLLM caps requests at 600 s by default; raise VLLM_MAX_AUDIO_DECODE_DURATION_S to 1800 for 30 minutes. Set max_num_seqs to 64 to fit 96 GB with the draft loaded.

What’s good

  • 2.3% error on Artificial Analysis, best open-weight model
  • Apache 2.0, commercial use allowed
  • Formatting instructions and key-term biasing
  • Second of 23 on VoiceCodeBench for entities

What’s not

  • 26.6B parameters; benchmarks ran on a 96 GB GPU
  • Seven languages, all European
  • No streaming; up to 30 minutes per file as tested
  • Custom plugin until vLLM and Transformers support lands; rare output loops

Best for

Enterprises in regulated industries that need near-commercial English and European-language transcription on their own GPUs.

Salesforce SOV-Transcribe-1 alternatives

Salesforce SOV-Transcribe-1: common questions

Is Salesforce SOV-Transcribe-1 worth it?
Open transcription that finally sits among the commercial leaders, though it wants a 96 GB GPU and speaks only seven languages. We rate it 4.3 out of 5.
How much does Salesforce SOV-Transcribe-1 cost?
Free under Apache 2.0; you pay for your own GPUs. No hosted price from Salesforce.
What are the downsides of Salesforce SOV-Transcribe-1?
26.6B parameters; benchmarks ran on a 96 GB GPU; seven languages, all European; no streaming; up to 30 minutes per file as tested; custom plugin until vLLM and Transformers support lands; rare output loops.
Who is Salesforce SOV-Transcribe-1 best for?
Enterprises in regulated industries that need near-commercial English and European-language transcription on their own GPUs.
How much does Salesforce SOV-Transcribe-1 cost for a team or in production?
No per-minute fee. With the draft model a 30-second clip takes about 0.22 s on one RTX PRO 6000 Blackwell; cost is that GPU's time.
Is Salesforce SOV-Transcribe-1 safe for company data?
Self-hosted under Apache 2.0, so audio stays on your servers. Training data and full training code are not part of the release.
What are the best Salesforce SOV-Transcribe-1 alternatives?
The closest alternatives we have reviewed are NVIDIA Nemotron 3.5 Transcribe, Voxtral 3 (Transcribe and TTS), Granite Speech 5.0 2B and ElevenLabs Scribe v3 and Scribe v3 Live.

Answers checked Nov 21, 2026.

Make Salesforce SOV-Transcribe-1? Show this rating on your siteSalesforce SOV-Transcribe-1 rated 4.3 out of 5 on ReviewRamble