Salesforce has released SOV-Transcribe-1, an open-weight speech-to-text model that now tops every other open model on the Artificial Analysis transcription leaderboard, with a 2.3% word error rate.
It is free under Apache 2.0, so companies can transcribe calls, meetings and voice notes on their own servers without sending audio to a cloud API. The catch is size: it needs a big GPU.
Below you get the specs, every benchmark number, hardware needs, setup with Transformers and vLLM, prompt tricks for names and keywords, comparisons, limits and fixes.
- What: a speech-to-text model from Salesforce’s sovereign AI team, released 12 November 2026.
- Build: Whisper large-v3 audio encoder plus a Gemma 4 26B A4B mixture-of-experts model. 26.6B total, 4.6B active parameters.
- Accuracy: AA-WER 2.28%, #1 open-weight model and #7-8 overall on Artificial Analysis.
- Languages: English (strongest), French, German, Spanish, Italian, Portuguese, Dutch.
- Licence: Apache 2.0. Commercial use allowed.
- Hardware: about 53 GB of BF16 weights. Tested on a 96 GB RTX PRO 6000 Blackwell. No quantized build yet.
- Extras: keyword biasing, style instructions (digits, no fillers), up to 30 minutes per file, non-streaming only.
What SOV-Transcribe-1 Is
SOV-Transcribe-1 is Salesforce’s first open-weight “speech-augmented language model”. Instead of a classic speech recogniser, it feeds audio into a large language model that writes the transcript.
The audio side is the well-known Whisper large-v3 encoder. The text side is Google’s open Gemma 4 26B A4B mixture-of-experts model, so only about 4.6B parameters work on each token.
Salesforce placed the release under its sovereign AI work, linked to Missionforce, its government and national-security effort. The goal is simple: let organisations run top-tier transcription on hardware they control.
Specs at a Glance
| Item | Detail |
|---|---|
| Developer | Salesforce Sovereign Intelligence (Missionforce Sovereign AI team) |
| Released | 12 November 2026, weights the same day |
| Architecture | Whisper large-v3 encoder + audio projector + Gemma 4 26B A4B MoE |
| Parameters | 26.6B total, 4.6B active (Hugging Face widget shows 27B) |
| Precision | BF16, MoE router in FP32 |
| Checkpoint | 16 safetensors shards, about 53 GB |
| Input | 16 kHz mono audio, optional text instructions and context |
| Max audio | Tested up to 30 minutes per file (vLLM default 600 s, raise to 1,800 s) |
| Mode | Non-streaming (whole file in, transcript out) |
| Languages | English, French, German, Spanish, Italian, Portuguese, Dutch |
| Speed-up model | SOV-Transcribe-1-DFlash draft for speculative decoding |
| Licence | Apache 2.0 |
Benchmarks: Every Published Number
The headline comes from Artificial Analysis, an independent leaderboard. Most of the other numbers are vendor-reported by Salesforce.
Artificial Analysis (independent)
| Model | AA-WER (lower is better) | Open weights |
|---|---|---|
| SOV-Transcribe-1 | 2.28% | Yes |
| Voxtral Small (Mistral) | 2.77% | Yes |
| Nemotron 3.5 Transcribe (NVIDIA) | 2.81% | Yes |
| Whisper Large v3 Turbo | 4.62% | Yes |
Its sub-scores are 1.79% on AA-AgentTalk, 1.50% on VoxPopuli and 4.06% on Earnings22. It ranked 7th of 44 systems at launch and 8th on the 13 November board.
Salesforce’s own results (vendor-reported)
| Test | SOV-Transcribe-1 | Rivals |
|---|---|---|
| VoiceCodeBench (full task success / entity recovery) | 76.3% / 93.4% | #2 of 23 systems, #1 open |
| AppTek call-centre, 14 English accents (avg WER) | 8.21% | Lowest on 8 accents incl. Indian, British, Australian |
| HALAS noisy earnings calls (WER) | 25.69% | Qwen3-ASR 1.7B 26.01%, Voxtral Small 27.35%, Nemotron 3.5 Transcribe 31.59%, Granite Speech 5.0 2B 34.51% |
| Preference-ASR, following style instructions (WER) | 5.05% | Granite Speech 5.0 2B 4.70%, Qwen3-ASR 1.7B 5.96%, Nemotron 3.5 Transcribe 6.16% |
| Contextual Earnings-22 keyword F1 | 71.9 no hints, 96.8 with clip keywords, 94.0 with full list | Best open model with full list |
Note: Indian English is one of the eight accents where it scored best in the call-centre test. That matters for Indian support teams, but the test is Salesforce’s own pick.
Speed and Latency
Salesforce measured on one RTX PRO 6000 Blackwell (96 GB). Times are per single request, with and without the DFlash draft model (vendor-reported).
| Audio length | Without draft | With DFlash draft |
|---|---|---|
| Up to 10 s | 0.16 s | 0.09 s |
| 10-30 s | 0.44 s | 0.15 s |
| 30-60 s | 0.88 s | 0.31 s |
| 1-5 min | 2.78 s | 1.26 s |
In batch serving, Salesforce reports up to 1,039 seconds of audio per second on that one card. In plain terms, one GPU can chew through about 17 hours of recordings every minute.
Hardware: Can Your GPU Run It?
This is where most home users stop. The BF16 weights alone are about 53 GB, before the memory needed for audio and transcripts.
- 96 GB card (RTX PRO 6000 Blackwell): the tested setup. Fits with the draft model if you set
max_num_seqs=64. - 80 GB card (H100, A100 80 GB): should fit the weights; keep batch sizes small.
- 48 GB card: does not fit in BF16 on its own.
- Gaming GPUs (16-32 GB): not possible until someone publishes a quantized build. None existed at launch.
Tip: on a normal PC, use Whisper locally or Nemotron 3.5 Transcribe (GGUF builds for 16-48 GB cards). Both are only slightly less accurate on the leaderboard.
How to Run It, Step by Step
Native support in Transformers and vLLM is still being added, so you use the model’s own code for now. Commands below follow the official model card.
Option 1: Transformers (quick test)
- Install the pinned versions.
pip install "transformers>=5.19,<5.23" torch librosa accelerate - Load the model with remote code. Use
AutoModelForMultimodalLM.from_pretrained(model_id, trust_remote_code=True, device_map="auto")with model idsalesforce-sovereign-intelligence/SOV-Transcribe-1, or theautomatic-speech-recognitionpipeline. - Feed 16 kHz mono audio. Load files with librosa at
sr=16000. - Generate. The card uses
model.generate(**inputs, max_new_tokens=8192)with greedy decoding.
Note: if your company blocks trust_remote_code, the model card includes a self-contained transcribe_hf function you can copy instead.
Option 2: vLLM server (recommended for real use)
- Download the plugin.
hf download salesforce-sovereign-intelligence/SOV-Transcribe-1 --include "vllm_plugin/*" --local-dir SOV-Transcribe-1 - Install it.
pip install ./SOV-Transcribe-1/vllm_plugin(this pins vLLM 0.32.0, so use a fresh virtual environment). - Start the server.
vllm serve salesforce-sovereign-intelligence/SOV-Transcribe-1 --max-model-len 32768 - Send a file. It exposes an OpenAI-style endpoint:
curl http://localhost:8000/v1/audio/transcriptions -F model=salesforce-sovereign-intelligence/SOV-Transcribe-1 -F file=@meeting.wav - Allow long files. Set
VLLM_MAX_AUDIO_DECODE_DURATION_S=1800before starting to accept 30-minute recordings.
Option 3: Add the speed-up draft model
Add a speculative config to the serve command that points at salesforce-sovereign-intelligence/SOV-Transcribe-1-DFlash with method dflash and num_speculative_tokens set to 15. Also set max_num_seqs to 64 on a 96 GB card, or vLLM’s default of 1,024 runs out of memory.
Prompts for Keywords and Style
Because a language model writes the transcript, you can steer it with a short instruction in the user message. Keep thinking turned off and use greedy decoding (temperature 0).
| Goal | Prompt from the model card |
|---|---|
| Verbatim | Transcribe the audio verbatim, outputting only spoken words in sequence. |
| Names and jargon | 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. |
| Known language | Transcribe the {language} audio into text, ensuring all punctuation marks are included. |
You can also ask for numbers as digits, cleaned-up text without “um” and “uh”, or a short description of the call (“a dental clinic booking call”) as context.
Tip: keyword lists are the biggest win. Salesforce’s own test jumped from 71.9 to 96.8 keyword F1 when product names were supplied.
SOV-Transcribe-1 vs Other Speech Models
| Model | AA-WER | Type | Runs on a gaming PC? | Best for |
|---|---|---|---|---|
| SOV-Transcribe-1 | 2.28% | Open, Apache 2.0 | No (53 GB BF16) | Most accurate self-hosted English and EU languages |
| Voxtral Small | 2.77% | Open | Large GPU | Open alternative from Mistral |
| Nemotron 3.5 Transcribe | 2.81% | Open | Yes, GGUF 16-48 GB | 40 languages, speakers and timestamps |
| Whisper Large v3 Turbo | 4.62% | Open, MIT | Yes, even CPU | Free, easy, many languages |
| ElevenLabs Scribe v3 | 1.9% | Cloud, $3.33 per 1,000 min | Not needed | Top accuracy without hardware |
| Microsoft MAI-Transcribe-2 | 2.0% | Cloud, $1.67 per 1,000 min | Not needed | Cheap, fast cloud batch |
AA-WER values are from Artificial Analysis. Cloud prices are list prices per 1,000 minutes shown on that leaderboard on 14 November 2026.
Limits, Licence and Privacy
- Seven languages only: no Hindi or other Asian languages. English is clearly the strongest.
- No streaming: it is not built for live captions.
- No speakers or timestamps: diarization and word-level timing are not documented.
- Repetition loops: it can occasionally repeat text until the token limit. The model card suggests a cleanup step.
- Early tooling: needs the vLLM 0.32.0 plugin or bundled modelling code until native support lands.
- Technical report: marked “coming soon”.
- Licence: Apache 2.0, so commercial use is fine. It is built on Gemma 4 and Whisper components; keep their notices.
Warning: transcribing calls or meetings records personal data. Tell people they are being recorded and follow your local rules (such as India’s DPDP Act or GDPR in Europe), even when everything runs on your own server.
Troubleshooting
- CUDA out of memory at startup: lower
max_num_seqs(64 on a 96 GB card), reduce--max-model-len, or drop the draft model. Our CUDA out of memory guide has more fixes. - “Model type not recognised” in Transformers: you are outside the 5.19-5.22 range or forgot
trust_remote_code=True. - vLLM rejects long files: set
VLLM_MAX_AUDIO_DECODE_DURATION_S=1800; the default cap is 600 seconds. - Transcript repeats the same line: a known issue. Split long files at pauses and add a step that removes repeated lines.
- Odd results or wrong speed: make sure audio is 16 kHz mono; resample with librosa or ffmpeg first.
- Speculative decoding errors in Transformers: it only handles one recording at a time there; use vLLM for batches.
My Take
This is a big moment for open speech models. A few months ago the gap between open and paid transcription was clear; now an Apache 2.0 model sits within a few tenths of a point of the best cloud services.
The trade-off is cost of ownership. At $1.67-$3.33 per 1,000 minutes for cloud APIs, you need a lot of audio, or strict data rules, before a 96 GB GPU pays off.
For most readers, the realistic path is to keep using Whisper or Nemotron at home and watch for a quantized SOV build. For banks, hospitals, call centres and government teams, this is now the open model to test first.
FAQ
Is SOV-Transcribe-1 free?
Yes. The weights are free under Apache 2.0, including for commercial use. You pay only for the hardware you run it on.
How much VRAM does SOV-Transcribe-1 need?
The BF16 weights are about 53 GB, so plan on an 80-96 GB GPU. Salesforce tested on a 96 GB RTX PRO 6000 Blackwell.
Can I run it on an RTX 4090 or 5090?
Not in its official form. A 24-32 GB card cannot hold the weights, and no quantized build was available at launch.
Does it support Hindi?
No. It supports English, French, German, Spanish, Italian, Portuguese and Dutch. For Hindi, use Whisper or Nemotron 3.5 Transcribe.
Is it better than Whisper?
On Artificial Analysis it scores 2.28% WER against 4.62% for Whisper Large v3 Turbo, so it makes about half the errors. Whisper is far lighter and supports many more languages.
Can it do live transcription?
No. It is non-streaming, so it works on finished recordings. Short clips return in well under a second on a large GPU.
Does it identify who is speaking?
Speaker labels and word timestamps are not documented. Pair it with a separate diarization tool if you need them.
Sources
- Salesforce, “Introducing SOV-Transcribe-1” Hugging Face blog: huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1
- Model card: huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1
- Artificial Analysis non-streaming speech-to-text leaderboard: artificialanalysis.ai/speech-to-text/non-streaming
- VoiceCodeBench leaderboard: besimple.ai/leaderboards/voice-code-bench
- TokenPost coverage (12 Nov 2026): tokenpost.com/news/technology/55846
- OrcaRouter analysis (hardware, sub-scores): orcarouter.ai/blog/sov-transcribe-1-salesforce-open-weights-speech-model
