Skip to main content

Salesforce SOV-Transcribe-1: Open-Weight Speech Model, Apache 2.0

SOV-Transcribe-1 is a speech-to-text model Salesforce released on 12 November 2026 with Apache 2.0 weights on Hugging Face. It has 26.6B parameters, about 4.6B active per token, and covers seven languages. Salesforce reports a 2.28% word error rate on an Artificial Analysis benchmark. It is self-hostable, but the official setup targets a large server GPU, not a laptop.

On this page

What Salesforce released

Salesforce's Sovereign Intelligence group published SOV-Transcribe-1 on Hugging Face on 12 November, calling it the company's first open-weight speech model. The weights are licensed under Apache 2.0, so they can be used commercially, fine-tuned, and run on your own machines.

It is a transcription model, not a voice assistant. It handles recorded audio, not live streams, takes 16 kHz input, and has been tested on recordings of up to 30 minutes. The model card lists seven languages: English, French, German, Spanish, Italian, Portuguese, and Dutch.

Salesforce pitches it at organisations that want to keep audio on infrastructure they control. That makes it relevant to anyone who transcribes meetings, calls, or interviews and does not want the recordings sent to a cloud service.

A language model that listens

Most speech recognisers are compact, single-purpose networks. SOV-Transcribe-1 is built differently: it joins the Whisper large-v3 audio encoder to Google's Gemma 4 26B A4B, a mixture-of-experts language model, through an audio projector. The card puts the total at 26.6B parameters with about 4.6B active per token, while the launch post rounds those to 26B and 4B.

Because a full language model does the writing, you can steer it with plain instructions. It transcribes verbatim by default, and you can ask for digits in place of spelled-out numbers, for filler words to be removed, or for different punctuation. You can also pass context, such as a list of names or product terms. In one Salesforce test on earnings calls, supplying each clip's keywords lifted keyword accuracy (F1) from 71.9 to 96.8.

The accuracy claims, with caveats

Salesforce says the model ranks first among open-weight models, and seventh overall, on Artificial Analysis's non-streaming speech-to-text benchmark as of 12 November, with a 2.28% word error rate. It lists Mistral's Voxtral Small at 2.77% and NVIDIA's Nemotron 3.5 Transcribe at 2.81%.

On harder audio the numbers are far less flattering for every model. On a set of real earnings-call recordings, Salesforce reports a 25.69% error rate, narrowly ahead of Qwen3-ASR 1.7B at 26.01%. On call-centre dialogues across 14 English accents, it reports 8.21% on average.

Treat all of this as vendor-reported. The comparison figures come from Salesforce's own runs, a technical report is still listed as coming soon, and the launch post states no limitations. A lead of a fraction of a point against a 1.7B model also raises a fair question about whether the extra size is worth it for your audio.

What it takes to run it yourself

This is where "open weights" and "runs locally" part ways. Salesforce's speed tests used a single NVIDIA RTX PRO 6000 Blackwell with 96 GB of memory, where a clip of 10 to 30 seconds took about 0.44 seconds. The card gives no minimum memory figure and does not mention CPU, phone, GGUF, or MLX support.

Setup is also early. Native support in vLLM and Transformers is still being upstreamed, so today you need a vLLM plugin or Transformers with remote code enabled. A separate 430M draft model, also Apache 2.0, speeds up single requests by about 2 to 3 times through speculative decoding.

Community 4-bit and 8-bit quantized versions appeared on Hugging Face within two days of launch. A 4-bit build of a 26.6B model should be far smaller than the full-precision weights, but those repositories are third-party, and nobody has published accuracy numbers for them yet.

Who should care

If you run a workstation or server GPU and need accurate, steerable transcription that never leaves your network, this is a strong new option under a permissive licence. If you want offline dictation on a phone or a thin laptop, smaller dedicated speech models remain the practical choice, at least until trustworthy quantized builds and wider runtime support arrive.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.