Is SOV-Transcribe-1 state of the art?Pending benchmarks.Announced Nov 12, 2026; no independent results yet.
Salesforce AI Research's open-weight speech-augmented language model for non-streaming automatic speech recognition. The official model card lists 26.6B total parameters (4.6B active), including a Whisper large-v3 audio encoder, an audio projector, and a Gemma 4 26B A4B mixture-of-experts decoder. It accepts up to 30 minutes of 16 kHz audio, optional natural-language instructions, and context biasing, and transcribes English, French, German, Spanish, Italian, Portuguese, and Dutch. Weights are ungated on Hugging Face at salesforce-sovereign-intelligence/SOV-Transcribe-1 under Apache 2.0, with a companion DFlash draft model for speculative decoding. No public self-serve API is stated.
The details so far
- Lab
- Salesforce
- Announced
- Nov 12, 2026
- API
- Not yet
- Open weights
- Published · Apache 2.0, ungated on Hugging Face at salesforce-sovereign-intelligence/SOV-Transcribe-1 (official card metadata license: apache-2.0; License section: Apache 2.0). Official launch blog: weights released on Hugging Face under Apache 2.0. A companion DFlash draft model for speculative decoding is at salesforce-sovereign-intelligence/SOV-Transcribe-1-DFlash; the official DFlash card says it cannot transcribe audio on its own
- Parameters
- 26.6B total (4.6B active), including the audio encoder
- Architecture
- Official model card: Whisper large-v3 audio encoder, audio projector, and Gemma 4 26B A4B mixture-of-experts decoder. Precision BF16, with the MoE router in FP32. Official card base models: openai/whisper-large-v3 and google/gemma-4-26B-A4B-it
- Audio input
- Official model card: up to 30 minutes (tested); 16 kHz audio with optional text instructions and context; output is transcript text
- Languages
- English, French, German, Spanish, Italian, Portuguese, and Dutch (official model card)
- Controls
- Official card: natural-language instructions for verbatim or cleaned-up transcripts, punctuation, capitalization, and number formatting; context biasing by supplying names, domain vocabulary, or a short topic description
- Inference
- Official card: vLLM (plugin in the repository for vLLM 0.32.0 until native support is released) and Hugging Face Transformers, with trust_remote_code modeling code in the repo
- Hosted access
- No public self-serve API or lab playground is stated on the official card or launch blog. Official blog: developers and enterprises can deploy the weights on infrastructure they control
Tags marked unsure are not stated on the lab's own pages.
What would let us call it
Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.
- Publicly announced with a source
- Usable by the public (API or weights)
- Listed on an independent board we track
- Enough results for a rank in a category
- Holds across two snapshots (crowns only)
What the lab claims
Self-reported. Shown for context; not used in any verdict.
Lab claim: self-reported official Artificial Analysis AA-WER Index 2.28% on the official model-card benchmark table and launch blog, which also report AA-AgentTalk 1.78, Earnings22-AA 4.06, and VoxPopuli-AA 1.50. The same official table and blog report Nemotron 3.5 Transcribe at 2.81% and Voxtral Small at 2.77% on the AA-WER Index.
Lab claim: self-reported official VoiceCodeBench (BeSimple) scores of Canonical Token/Entity Match 93.39, Task Success Rate 76.33, and WER 5.01 (official model-card table). The official launch blog rounds these to 93.4% of entities recovered and 76.3% of recordings passed.
Lab claim: self-reported 8.21% average word error rate on AppTek Call-Center Dialogues across 14 English accents, and 25.69% word error rate on HALAS, from the official model-card table and launch blog.
Lab claim: self-reported Preference-ASR preference-aware word error rate of 5.05% with instructions and 5.72% with no instructions (official model-card table and launch blog).
Lab claim: self-reported Contextual Earnings-22 keyword F1 of 71.9 with no keywords, 96.8 with the clip's keywords, and 94.0 with the full-call keyword list (official launch blog). The official table also reports word error rates of 12.42, 9.73, and 9.98 on those three conditions.
Lab claim: self-reported official model-card table, Open ASR English public-set mean word error rate 5.44 and Open ASR multilingual mean 4.88.
Sources
LABIntroducing Salesforce SOV-Transcribe-1 (official launch blog)huggingface.co/blog/salesforce-sovereign-intelligence/sov-transcribe-1 · Nov 12, 2026REPOSOV-Transcribe-1 on Hugging Facehuggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1 · Nov 12, 2026LABSalesforce AI Research announcement on Xx.com/SFResearch/status/2121092946254823642 · Nov 12, 2026LABMarc Benioff announcement on Xx.com/Benioff/status/2121077110488109147 · Nov 12, 2026Tracking continues until an independent board lists it.