Salesforce launches SOV-Transcribe-1, open speech transcription model topping the open ranking

Salesforce launched SOV-Transcribe-1 on November 12, 2026, its first open-weight speech transcription model. The model turns audio into text, runs on the user's own infrastructure and ships under the Apache 2.0 license, available for download on Hugging Face.

What the model does

SOV-Transcribe-1 is a non-streaming automatic speech recognition (ASR) model: it takes audio files of up to 30 minutes and returns the transcript. Among the features highlighted by Salesforce are:

  • Entity recognition: it captures names, phone numbers and email addresses, details that matter in dictation and voice applications;
  • Robustness to accents and noise: it handles diverse English accents and challenging audio conditions, including background noise;
  • Style control by instruction: it accepts natural-language commands to choose between verbatim or cleaned-up transcripts and to control punctuation, capitalization and number formatting;
  • Context: it accepts names, domain vocabulary and a short description of the topic to recognize difficult words and technical terms.

Benchmark results

According to the model card on Hugging Face, SOV-Transcribe-1 ranks first among open-weight models for non-streaming speech-to-text on Artificial Analysis' AA-WER index, with a word error rate of 2.28%, and seventh overall. On VoiceCodeBench, it places second out of 23 systems for entity accuracy and task success.

Among open competitors, Mistral AI's Voxtral Small scores 2.77% and NVIDIA's Nemotron 3.5 Transcribe scores 2.81% on the same index. Geekiki already covered the launch of Nemotron 3.5 Transcribe in late October.

Architecture and how to use it

The model combines a Whisper large-v3 audio encoder with a projector and a Gemma 4 26B A4B mixture-of-experts decoder. It has 26.6 billion parameters in total, with 4.6 billion active at each step, in BF16 precision. The supported languages are English, French, German, Spanish, Italian, Portuguese and Dutch.

The model works with vLLM and Hugging Face Transformers, and a companion model, SOV-Transcribe-1-DFlash, proposes 15 tokens at a time to speed up decoding by about 2 to 3 times. Native support in both libraries is still being upstreamed: for now, a vLLM plugin or the modeling code shipped with the repository is required.

Who it is for — and its limits

SOV-Transcribe-1 is part of Salesforce's Sovereign AI initiative, aimed at public agencies and companies that cannot send audio recordings to third-party servers. The idea is for each organization to run the model on its own infrastructure and keep control of its data.

There are important caveats. The model is non-streaming by design, so it was not built for real-time transcription. The BF16 checkpoint is about 53 GB, which demands powerful hardware to run locally. And the benchmark numbers are reported by Salesforce itself: two of the test sets used in the index are not public, so they cannot be independently reproduced.

Sources

Featured image: official SOV-Transcribe-1 banner, published by Salesforce on the model's Hugging Face page.

Transparency: this article was produced with the help of artificial intelligence, cross-checked against public posts and internet sources. Check the original sources to confirm the facts before sharing.

Por Muse