Introducing Salesforce SOV-Transcribe-1: #1 Open-Weight Model Built for Real-World Speech

Today, we’re introducing a new open-weight speech model that intelligently understands what you say, from the email address you dictate to the way you want the transcript formatted.
SOV-Transcribe-1 — our first open-weight speech-augmented language model, with 4 billion active parameters (26 billion total), built for non-streaming automatic speech recognition (ASR). It offers:
For developers and enterprises, this open-weight model (License: Apache 2.0) delivers the building blocks for sovereign, enterprise-ready voice experiences that you retain control over.
SOV-Transcribe-1 ranks #1 among open-weight models for non-streaming speech-to-text on Artificial Analysis (#7 overall), and #2 of 23 systems on VoiceCodeBench. Model inclusion and rankings as of November 12, 2026.
With 4B active parameters, SOV-Transcribe-1 pushes the Pareto frontier of Artificial Analysis word error rate (AA-WER) vs. active parameters among open-weight models. At 2.28% WER, it beats NVIDIA’s Nemotron 3.5 Transcribe (2.81%) with the same number of active parameters, and Mistral’s Voxtral Small (2.77%) with 6x fewer (4B vs. 24B). Other open-weight models of a similar size sit at around 4% WER.
Not all words are created equal. For example, in a voice agent, a handful of words carry the task itself: an email address, an account number, a command. A transcript that looks mostly correct can still hide a critical error (a dropped underscore, a swapped digit) that surfaces downstream as an incorrect tool argument, an invalid database field, or a completely different meaning.
VoiceCodeBench (BeSimple's Private Leaderboard) measures whether a speech-to-text system captures the key terms that critical workflows depend on, from phone numbers and emails to URLs and code symbols. Its Task Success Rate (TSR) passes a recording only if every target entity in it is correct, while Canonical Token/Entity Match (CTEM) measures the share of entities recovered. SOV-Transcribe-1 passes 76.3% of recordings and recovers 93.4% of entities, ranking #1 among open-weight models and #2 among 23 systems overall on both measures.
Even a single language is spoken in many different ways. A transcription model built for the enterprise has to work equally well for a customer speaking English in Glasgow, Dublin, Singapore, or Mumbai.
On AppTek Call-Center Dialogues, a long-form benchmark of agent–customer calls spanning 14 English accents, SOV-Transcribe-1 achieves an 8.21% average word error rate. It has the lowest WER on 8 of the 14 accents: Australian, British, Scottish, Welsh, Irish, Indian, Mexican, and Singaporean English.
Speech rarely happens in clean, studio-quality environments. More often, people are speaking in a busy café or on a crowded subway, where the costliest failure is missing a key word or hallucinating one that was never said.
To measure this, we evaluate on HALAS, a human-annotated benchmark of real earnings-call audio that models are most likely to hallucinate on. SOV-Transcribe-1 achieves a 25.69% word error rate, lower than all other open-weight models, including Alibaba’s Qwen3-ASR 1.7B (26.01%), Mistral’s Voxtral Small 24B (27.35%), NVIDIA’s Nemotron 3.5 Transcribe (31.59%), and IBM’s Granite Speech 5.0 2B (34.51%).
Different applications need different transcripts. A deposition needs every word verbatim, including fillers, while a dictated message needs clean, formatted text. SOV-Transcribe-1 is instruction-controlled: it transcribes verbatim by default and adapts its formatting to natural-language instructions, such as writing numbers as digits or removing disfluencies.
Preference-ASR measures style steering: it scores whether a transcript follows a requested format across numbers, entities, disfluencies, and casing, first with no instructions and then with the preference given as an instruction. With instructions, SOV-Transcribe-1 achieves a 5.05% preference-aware word error rate, second only to IBM’s Granite Speech 5.0 2B (4.70%) and ahead of Alibaba’s Qwen3-ASR 1.7B (5.96%) and NVIDIA’s Nemotron 3.5 Transcribe (6.16%).
A customer’s name on a sales call or a drug name in a clinical note can be difficult to transcribe, because they’re often rare or sound like other words. Give SOV-Transcribe-1 keywords to boost or a short description of the audio, and it resolves those words toward the terms that fit.
Contextual Earnings-22 tests this on 15-second clips of real earnings calls, each built around the name of a person, company, or product. Each clip is transcribed three ways: with no keywords, with only the keywords spoken in the clip, and with every keyword from the full hour-long call, most of them distractors. Given just the clip’s keywords, SOV-Transcribe-1 raises its keyword F1, a measure of how accurately those names are transcribed, from 71.9 to 96.8. Given the full list, it still scores 94.0, the highest of the open-weight models.
Full results for SOV-Transcribe-1, NVIDIA’s Nemotron 3.5 Transcribe, Alibaba’s Qwen3-ASR 1.7B, IBM’s Granite Speech 5.0 2B, and Mistral’s Voxtral Small 24B, across the benchmarks above plus Open ASR (English and multilingual) and ContextASR-Bench.
With SOV-Transcribe-1, developers and enterprises can build voice agents, dictation tools, and call transcription while keeping critical intelligence on infrastructure they control. We’re releasing the weights today on Hugging Face under the Apache 2.0 license.
Missionforce Sovereign AI: Tanay Dixit, Daniel Lee, Ansh Moraje, Aman Gupta, Young Mo Kang, Alexey Dyakov, Kah Mun Chia, John Emmons
· Sign up or log in to comment