The Pulse
Salesforce releases open-weight speech transcription model
Salesforce has released SOV-Transcribe-1 under an Apache 2.0 license for non-streaming speech recognition. The company says the model leads open-weight systems on an Artificial Analysis transcription benchmark.

Salesforce has released an open-weight speech model for non-streaming transcription under an Apache 2.0 license. In a Nov. 12 post on Hugging Face, the company's Sovereign Intelligence team said SOV-Transcribe-1 supports multiple languages and lets developers use instructions to control transcript formatting.
The company says it ranked first among open-weight models and seventh overall for non-streaming speech-to-text on Artificial Analysis, with rankings current as of Nov. 12. The weights are available on Hugging Face.
Model and accuracy
SOV-Transcribe-1 has 4 billion active parameters and 26 billion total parameters. It is designed for automatic speech recognition and can be used in applications including voice agents, dictation tools and call transcription on infrastructure controlled by developers and enterprises.
Salesforce reported a 2.28% word error rate for the model on the Artificial Analysis comparison, against 2.81% for NVIDIA's Nemotron 3.5 Transcribe, which has the same number of active parameters. The company put Mistral's Voxtral Small at 2.77%, with 24 billion active parameters.
Names, numbers and accents
The release puts particular emphasis on recognizing details such as phone numbers, email addresses, URLs and code symbols. VoiceCodeBench measures both whether a recording has every target entity correct and the share of individual entities recovered.
The model correctly captured every target entity in 76.3% of recordings and recovered 93.4% of entities, according to Salesforce. The company said those scores placed it first among open-weight models and second among 23 systems overall on both measures.
On the AppTek Call-Center Dialogues benchmark of agent-customer calls spanning 14 English accents, the model recorded an 8.21% average word error rate, the company reported. It had the lowest error rate on 8 of the 14 accents, including Scottish, Indian and Singaporean English.
Salesforce also reported a 25.69% word error rate on HALAS, a human-annotated benchmark using real earnings-call audio. It said the result was lower than those of the other open-weight models it compared, including Alibaba's Qwen3-ASR 1.7B at 26.01%.
Formatting and context
The model produces verbatim transcripts by default, but natural-language instructions can ask it to remove disfluencies or write numbers as digits. Preference-ASR tests whether transcripts follow requested formats for numbers, entities, disfluencies and casing.
With instructions, the model achieved a 5.05% preference-aware word error rate on that test, Salesforce said. That placed it behind IBM's Granite Speech 5.0 2B, at 4.70%, in the company's comparison.
Developers can also supply keywords or a short audio description to guide transcription of difficult terms. In the Contextual Earnings-22 test, which uses 15-second clips from real earnings calls, Salesforce compared transcription without keywords against runs supplied with either the clip's keywords or the full call's keyword list.
Its keyword F1 score rose from 71.9 without keywords to 96.8 with the clip's keywords, the company reported. Given the full call list, which included distractors, it scored 94.0; Salesforce said that was the highest result among open-weight models.