Not tested by us yet. SOV-Transcribe-1 is an open-weights release that needs a 96 GB-class GPU for the setup Salesforce documents, and our harness calls hosted APIs only. Nothing hosts it yet. Everything below comes from the model card on Hugging Face, Salesforce's launch post and the Artificial Analysis leaderboard. We'll update this post when a hosted endpoint lets us run it on our own audio.
SOV-Transcribe-1 is a speech-to-text model that transcribes recorded audio in English, French, German, Spanish, Italian, Portuguese and Dutch, and can follow plain-language instructions about style and vocabulary. If you want a self-hosted transcriber with strong accuracy on names, phone numbers and emails, it is the new open-weights model to benchmark. If you need live streaming transcription, it is not built for that. Salesforce reports an Artificial Analysis word error rate of 2.28%, the best among open-weights models there.
| SOV-Transcribe-1 | |
|---|---|
| Maker | Salesforce (Sovereign Intelligence team) |
| Released | 2026-11-12 |
| Architecture | Whisper large-v3 audio encoder, audio projector, Gemma 4 26B A4B mixture-of-experts decoder |
| Size | 26.6B total, 4.6B active, including the audio encoder |
| Languages | English, French, German, Spanish, Italian, Portuguese, Dutch |
| Input | 16 kHz audio, tested up to 30 minutes, optional text instructions and context |
| Mode | Non-streaming only |
| License | Apache 2.0 |
| Price | Free weights, you pay for your own GPU |
What is new
Most open speech models are plain encoder-decoder transcribers. This one is a speech-augmented language model: audio goes through a Whisper encoder into a Gemma 4 language model, so you can tell it what you want in words. The model card lists verbatim or cleaned-up output, punctuation and number formatting, and context biasing with a list of names or a short description of the topic.
The card's sample prompts are short. The default is "Transcribe the audio verbatim, outputting only spoken words in sequence." To bias toward your vocabulary you add a comma-separated list of terms the recording may contain, and to fix the language you name it in the instruction.
The numbers, and who produced them
Salesforce's launch post reports these scores. They are vendor-reported, and we haven't reproduced any of them.
| Benchmark | SOV-Transcribe-1 | Compared with (as reported) |
|---|---|---|
| Artificial Analysis WER | 2.28% | Voxtral Small 2.77%, Nemotron 3.5 Transcribe 2.81%, other similar open models around 4% |
| HALAS (earnings-call audio) | 25.69% | Qwen3-ASR 1.7B 26.01%, Voxtral Small 27.35%, Nemotron 3.5 Transcribe 31.59% |
| AppTek call-center dialogues | 8.21% average over 14 English accents | Lowest WER on 8 of 14 accents |
| Preference-ASR (style steering) | 5.05% | Granite Speech 5.0 2B 4.70% ranks first |
| VoiceCodeBench task success | 76.3% | #2 of 23 systems |
Two things stand out. First, the Artificial Analysis lead over Voxtral Small is 0.49 points, and Salesforce notes Voxtral Small is a 24B-parameter model. Second, the model does not win everywhere: on Preference-ASR a 2B Granite model beats it, and the HALAS gap to Qwen3-ASR 1.7B is under half a point.
Check the table against the public leaderboard before you quote it. The Artificial Analysis table lists hosted models with prices, and it shows Nemotron 3.5 Transcribe at 3.4% on DeepInfra, while its FAQ and Salesforce both give about 2.8%. Hosting and settings change the score, which is a reason to measure your own audio. For reference, that table also lists ElevenLabs Scribe v3 at 1.9% ($3.33 per 1,000 minutes) and Microsoft's MAI-Transcribe-2 at 2.0% ($1.67), both ahead of this model and both proprietary.
Running it
Speed is the strongest practical claim. On one RTX PRO 6000 Blackwell with vLLM 0.32.0, the card reports up to 1,039 seconds of audio per second and a 30-second clip in 0.22 seconds, using a separate draft model (SOV-Transcribe-1-DFlash) for speculative decoding. Without the draft, the card lists 0.44 seconds for 10 to 30 second clips. A third-party write-up puts the BF16 weights alone at roughly 53 GB.
Native Transformers and vLLM support is still being upstreamed. Today you install a vLLM plugin from the repo or load the model with trust_remote_code=True. The vLLM server exposes an OpenAI-compatible /v1/audio/transcriptions endpoint, which makes it a drop-in for code written against that API.
Limits the card admits
- Non-streaming only, so no live captions.
- Rare repetition loops, where the model repeats a character or phrase until the token limit. The card recommends post-processing cleanup.
- Long recordings should be split at speech boundaries and checked for truncated or repeated output. The default vLLM endpoint accepts 600 seconds per request.
- Training data is not disclosed and the technical report is listed as "coming soon".
- Seven languages, all European. Nothing for Hindi, Japanese or Arabic, for example.
- Greedy decoding with thinking disabled is the recommended setting.
Who should care
Teams that can't send audio to a third-party API, such as call recordings with customer data, get a permissively licensed model near the top of a public leaderboard. Teams with a single data-center GPU and a backlog of earnings calls or support recordings are the clearest fit. Everyone else should compare it against hosted options on price per 1,000 minutes, where Whisper Large v3 and Nemotron 3.5 Transcribe are listed at $1.15 to $4.23 on the leaderboard.
For another sovereignty-minded teardown, see our Protegrity-PII write-up, and for the wider picture read which LLMs suit privacy-first apps. If you transcribe calls for agents, our teardown of Retell Tell-1 covers the voice-agent side.