Salesforce releases open transcription model ranked first among open systems
Salesforce published SOV-Transcribe-1, an open-weights speech recognition model, on Hugging Face under Apache 2.0.
· Last verified: 15 Nov 2026 (Hugging Face, Artificial Analysis, besimple AI)

Summary
- Salesforce published SOV-Transcribe-1, an open-weights speech recognition model, on Hugging Face under Apache 2.0.
- The model card ranks it first among open non-streaming transcription systems and seventh overall on Artificial Analysis.
- It covers seven European languages; Arabic is absent, leaving that market open to rivals.
The latest
An open-weights speech-to-text model that runs on a single GPU and carries a commercial-use license is now public: Salesforce Sovereign Intelligence posted SOV-Transcribe-1 to Hugging Face under Apache 2.0. The model card places it first among open non-streaming transcription systems on Artificial Analysis, and second out of 23 systems on besimple AI's VoiceCodeBench.
Details
- The architecture: The model pairs a Whisper large-v3 audio encoder with an audio projection layer and a Gemma 4 26B A4B mixture-of-experts language model, according to the model card. Total parameters reach 26.6 billion, of which 4.6 billion are active including the audio encoder. Precision is BF16, with the expert router at FP32.
- The benchmark claim: Artificial Analysis said on its official X account that SOV-Transcribe-1 is the top open speech-to-text model, scoring 2.3% on the AA-WER index with 4.6 billion active parameters, and ranking 7th of 44 models including closed systems. The model card cites rankings as of 12 November 2026.
- The speed numbers: On a single RTX PRO 6000 Blackwell Server Edition with 96 gigabytes of memory, the card reports throughput up to 1,039 seconds of audio per second, with a 30-second clip returned in 0.22 seconds. Maximum tested audio length is 30 minutes.
- Speculative decoding: A separate DFlash draft model, in its own repository, cuts single-request latency two to three times: 0.09 seconds instead of 0.16 for clips up to 10 seconds, 0.15 instead of 0.44 for 10-30 seconds, 0.31 instead of 0.88 for 30-60 seconds, and 1.26 instead of 2.78 for one to five minutes.
- The language gap: Coverage is limited to seven languages — English, French, German, Spanish, Italian, Portuguese and Dutch. Arabic is not among them, meaning enterprises working in Arabic gain nothing directly from this release and the competitive field in Arabic transcription stays unsettled.
- What it does: Listed tasks include verbatim or cleaned transcription, multilingual recognition, style control and contextual transcription. The card highlights entity accuracy for names, phone numbers and email addresses, robustness to accents and background noise, and contextual biasing using supplied names, vocabulary or a topic description.
- How to run it: Deployment requires Transformers version 5.19 or higher and below 5.23 with trust_remote_code, or vLLM through a plugin that pins version 0.32.0. An OpenAI-compatible endpoint at /v1/audio/transcriptions accepts up to 600 seconds per request by default, extendable to 1,800 via an environment variable.
- The caveats: Salesforce did not publish official minimum hardware requirements; the figures were measured on a 96-gigabyte GPU. The card warns that vLLM's default max_num_seqs of 1024 exhausts memory when the draft model is loaded, and recommends 64. Native support in vLLM and Transformers is described as still being upstreamed.
- Usage guidance: The card recommends stating the output style explicitly, including number formatting, supplying focused context, feeding mono 16 kilohertz audio, keeping reasoning disabled, and using greedy decoding with temperature at 0. It also advises checking outputs for truncation or repetition loops.
Background
Whisper, the encoder reused here, set the reference point for open speech recognition. SOV-Transcribe-1 keeps that encoder but routes its output into a mixture-of-experts language model, which is why only 4.6 billion of 26.6 billion parameters run per token.
Between the lines
The combination that matters commercially is Apache 2.0 licensing, single-GPU operation and a top-ranked open score at 4.6 billion active parameters. Together they let enterprises keep audio on their own infrastructure rather than sending it to a closed API. The seven-language limit defines where that applies: for Arabic-language workloads, the proprietary providers face no new open challenger from this release.
What's next
Watch for native support landing in vLLM and Transformers, which would remove the trust_remote_code and plugin requirements, and for independent VoiceCodeBench and Artificial Analysis results confirming the posted rankings.