Salesforce releases SOV-Transcribe-1, a top-ranked open-weight speech model

Salesforce published an Apache 2.0 open-weight speech transcription model that it says ranks first among open-weight models on several benchmarks.

Salesforce released SOV-Transcribe-1, an open-weight automatic speech recognition model, on November 12, 2026, saying the model ranks first among open-weight systems on several speech benchmarks. The announcement came in a blog post introducing SOV-Transcribe-1 published jointly by Hugging Face and Salesforce.

What changed

SOV-Transcribe-1 is a non-streaming ASR / speech-augmented language model with 4 billion active parameters and 26 billion total parameters.
The model is released under the Apache 2.0 license as open weights on Hugging Face.
On the Artificial Analysis benchmark it reaches a word error rate of 2.28%, ranking #1 among open-weight models and #7 overall.
On VoiceCodeBench it achieves a 76.3% task success rate, #1 among open-weight models and #2 overall, with 93.4% entity match.
On the AppTek Call-Center benchmark, covering 14 accents, the average word error rate is 8.21%.
The model supports multilingual transcription, entity recognition, accent robustness, instruction-controlled style steering, and domain context-biasing.

What it means for people building with AI

For teams building voice features, an open-weight model with reported state-of-the-art accuracy among open models gives another option beyond closed APIs, under a permissive Apache 2.0 license that allows commercial use and modification. The benchmark spread also matters: Salesforce reports a HALAS real-world audio word error rate of 25.69%, much higher than the curated Artificial Analysis score, and a contextual Earnings-22 keyword F1 score of 96.8% when relevant keywords are supplied, meaning actual accuracy in production settings depends heavily on whether domain context can be provided to the model.

Availability

SOV-Transcribe-1 is available now as open weights on Hugging Face under the Apache 2.0 license. Its release lands the same week Hugging Face introduced its decision-model category, extending the open-weight race beyond text language models into audio.

Sources

Introducing Salesforce SOV-Transcribe-1 (Hugging Face / Salesforce, November 12, 2026, official)

Written with AI assistance, strictly from the sources listed above. Before it went live, automated checks verified its dates, sources and every language edition.