Salesforce released SOV-Transcribe-1 on November 12, 2026, an open-weight 26.6-billion-parameter speech recognition model distributed under the permissive Apache 2.0 license. Combining a Whisper audio encoder with a Mixture-of-Experts language backbone, the model enables local, non-streaming transcription with natural-language steering. For university researchers and research administration offices, local execution eliminates third-party cloud data exposure during interview transcription, establishing a secure alternative for institutional compliance.
Architecture and controllable transcription features
Developed by Salesforce Sovereign Intelligence, the system pairs an audio encoder based on Whisper large-v3 and an audio projector with a Gemma-based Mixture-of-Experts (MoE) text decoder (Gemma 4 26B A4B MoE). While the total parameter count is 26.6 billion, only 4.6 billion parameters are activated per token, keeping inference requirements manageable on enterprise hardware. Model weights are configured in BF16 precision, with the MoE routing layer maintained in FP32. The model processes 16 kHz audio files up to 30 minutes in length and supports seven languages: English, French, German, Spanish, Italian, Portuguese, and Dutch, as documented in AI and ML research outputs.
Unlike conventional automatic speech recognition (ASR) pipelines that produce fixed text, SOV-Transcribe-1 accepts natural-language prompting to direct transcription formatting. Researchers can steer output style between verbatim transcriptions that retain conversational disfluencies and cleaned-up, publication-ready prose. The architecture also supports context biasing, allowing users to supply target keyword glossaries or brief topic summaries to resolve specialized terminology, participant names, and technical vocabulary without model fine-tuning.
Reported benchmark performance and local deployment
Salesforce reports that SOV-Transcribe-1 achieved a 2.28% word error rate (WER) on the Artificial Analysis non-streaming benchmark, ranking first among open-weight speech models and seventh overall including commercial proprietary engines. In comparative evaluations published by the developer, the model surpassed NVIDIA Nemotron 3.5 Transcribe (2.81% WER) and Mistral Voxtral Small (2.77% WER). On the VoiceCodeBench evaluation for critical entity recognition, Salesforce reported a 76.3% Task Success Rate and a 93.4% Canonical Token/Entity Match score across phone numbers, identifiers, and technical symbols. Across 14 regional English accents in AppTek call-center evaluations, the model achieved an 8.21% average WER, leading eight regional subsets.
Model weights and code are hosted publicly on Hugging Face. The model can be served using Hugging Face Transformers or vLLM via a repository plugin, with speculative decoding supported through an accompanying DFlash draft model to reduce inference latency.
Implications for qualitative research and institutional compliance
The availability of a high-accuracy, open-weight speech model directly addresses a recurring data governance bottleneck in academic institutions. In qualitative studies involving participant interviews, fieldwork audio, or clinical observations, transcribing spoken recordings is foundational to qualitative data analysis. However, institutional policies and human-subjects research determination protocols governed by Institutional Review Boards (IRBs) frequently prohibit researchers from uploading identifiable voice recordings to commercial cloud-hosted transcription APIs.
Voice recordings contain biometrically unique audio signatures and frequently touch upon sensitive personal disclosures or protected health data. Deploying SOV-Transcribe-1 within university-managed data centers or isolated departmental workstations fulfills institutional data stewardship mandates. Research teams maintain complete custody of primary audio, eliminating third-party cloud data retention risks while benefiting from specialized promptable transcription controls.
Sources
- Hugging Face Blog, Introducing Salesforce SOV-Transcribe-1: #1 Open-Weight Model Built for Real-World Speech, November 12, 2026. Checked 2026-11-14.
- Hugging Face Model Card, salesforce-sovereign-intelligence/SOV-Transcribe-1, November 12, 2026. Checked 2026-11-14.
- Superpower Daily, Salesforce releases open-weight transcription AI, claiming a benchmark lead, November 13, 2026. Checked 2026-11-14.
