Introduction

Salesforce (its Missionforce Sovereign AI team) released SOV-Transcribe-1 on November 12. It's an open-weight speech recognition (ASR) model under Apache 2.0: a Whisper large-v3 audio encoder feeding a Gemma 4 26B-A4B mixture-of-experts decoder, with 26.6B total and 4.6B active parameters. It does non-streaming transcription in English plus French, German, Spanish, Italian, Portuguese and Dutch. You can run it today from Hugging Face with a vLLM plugin (including an OpenAI-compatible /v1/audio/transcriptions server) or with Transformers.

The thing that sets it apart from a plain Whisper-style model is that it's instruction-driven. You tell it, in plain language, whether you want a verbatim transcript or a cleaned-up one and how to format numbers, and you can hand it a list of names and jargon to bias toward. On the same day, Artificial Analysis put it at #1 among open-weight speech-to-text models.

Sources: Salesforce's launch post on Hugging Face (published November 12, 2026), the SOV-Transcribe-1 model card, the DFlash draft model, and Artificial Analysis's speech-to-text leaderboard with its

.

What Shipped

  • Name and maker: SOV-Transcribe-1, from Salesforce's Missionforce Sovereign AI team. It's published under the salesforce-sovereign-intelligence org on Hugging Face.
  • Modality: speech-to-text only, non-streaming. It takes 16 kHz audio up to 30 minutes long (the tested maximum) and returns transcript text.
  • Architecture: a Whisper large-v3 encoder, an audio projector, and a Gemma 4 26B-A4B MoE decoder. That adds up to 26.6B total parameters, with 4.6B active per token including the encoder.
  • Languages: English, French, German, Spanish, Italian, Portuguese and Dutch.
  • License: Apache 2.0, ungated. The BF16 weights come in 16 safetensors shards, roughly 53 GB by our arithmetic (26.6B × 2 bytes).
  • Speed add-on: a separate DFlash draft model for speculative decoding proposes 15 tokens at a time. Salesforce says it gives about a 2–3× speedup for single requests, with accuracy within run-to-run noise.
  • Not included: streaming, and a technical report (marked "coming soon"). The model card doesn't list a public hosted API.

How to Run It

  • vLLM (recommended): native support is still being upstreamed. For now, download the repo's vllm_plugin/ folder and pip install it, which pins vLLM 0.32.0. Then vllm serve salesforce-sovereign-intelligence/SOV-Transcribe-1 --max-model-len 32768 gives you an OpenAI-compatible transcription endpoint you can hit with curl -F file=@meeting.wav.
  • Long files on the server: the endpoint caps requests at 600 seconds unless you set VLLM_MAX_AUDIO_DECODE_DURATION_S=1800 for the full 30 minutes.
  • Transformers: use transformers>=5.19,<5.23 with trust_remote_code=True and processor.apply_transcription_request(audio, prompt=...). The card also includes a self-contained function for environments that block remote code.
  • Hardware: Salesforce benchmarks on a single 96 GB RTX PRO 6000 Blackwell. With the draft model, it advises max_num_seqs=64, because vLLM's default of 1,024 sequences runs out of memory.
  • Decoding: use greedy settings (temperature=0) and keep thinking disabled. Rarely, the model loops on a phrase until it hits the token limit, and the card suggests a simple cleanup pass adapted from the Qwen3-ASR toolkit.

Prompts That Steer It

Put the instruction in the user message. These are from the card's prompt guide:

  • Verbatim: Transcribe the audio verbatim, outputting only spoken words in sequence. This is also the server default.
  • Keyword biasing: This audio may contain the following words or phrases: {keywords}. Transcribe the audio into text.
  • Auto language: Transcribe the audio verbatim in the language's standard writing system, using punctuation and capitalization. Output only the transcript.
  • Language hint: Transcribe the {language} audio into text, ensuring all punctuation marks are included.
  • You can also ask in plain language for digits instead of spelled-out numbers, or for disfluencies to be removed.

Numbers

Independent: Artificial Analysis runs these on its own.

  • AA-WER Index: 2.3% WER, the best open-weight result and #7 of 44 models overall including proprietary ones. Mistral's Voxtral Small and NVIDIA's Nemotron 3.5 Transcribe are next among open models, both at 2.8%.
  • Per dataset: 4.1% on Earnings22 (Voxtral Small 4.8%), 1.5% on VoxPopuli (1.8%), and 1.8% on AA's AgentTalk voice-agent set.
  • Speed: AA measured a 60× speed factor on Salesforce's own hosted endpoint, which works out to about one second per minute of audio. That endpoint isn't a public product listed on the card.

Self-reported: these come from Salesforce's launch post and model card.

  • VoiceCodeBench (BeSimple): 76.3% of recordings had every entity correct, and 93.4% of entities were recovered (phone numbers, emails, URLs). That's #2 of 23 systems and #1 among open-weight models.
  • AppTek call-center dialogues: 8.21% average WER across 14 English accents. Salesforce says it had the lowest WER on 8 of them, including Scottish, Indian and Singaporean English.
  • HALAS (hallucination-prone earnings-call audio): 25.69% WER, against 26.01% for Qwen3-ASR 1.7B and 27.35% for Voxtral Small.
  • Context biasing (Contextual Earnings-22): keyword F1 rose from 71.9 to 96.8 when given the clip's keywords, and stayed at 94.0 when given a long list that was mostly distractors.
  • Style steering (Preference-ASR, instructed): 5.05% preference-aware WER, behind IBM Granite Speech 5.0 2B at 4.70%.
  • Throughput: one RTX PRO 6000 handles up to 1,039 seconds of audio per second, and a 30-second clip comes back in 0.22 seconds (BF16 with DFlash, vLLM dev build).

Caveats

  • It's batch-only. For live captions or voice-agent turn-taking you'll still need a streaming model.
  • At 26.6B total parameters, the "4.6B active" figure helps speed but not memory. Plan for a big single GPU, not a laptop.
  • vLLM and Transformers support is plugin and remote-code only until it's merged upstream. The plugin pins vLLM 0.32.0.
  • The technical report isn't out yet, and everything except the Artificial Analysis figures is vendor-run.

Conclusion

SOV-Transcribe-1 is an Apache-2.0 open-weight ASR model you can self-host today. It's the current #1 open-weight model on Artificial Analysis, and its best tricks are taking plain-language formatting instructions and keyword lists for names and jargon. If you transcribe calls, meetings or dictation on your own hardware, run it against Nemotron 3.5 Transcribe and Voxtral Small on your own audio, especially recordings full of numbers and emails.

—Anabel ♡