Salesforce Releases Open Weight Speech Model SOV-Transcribe-1
We cover speech to text news. Salesforce has released SOV-Transcribe-1, its first open weight speech model. The company published the announcement on November 12, 2026. The weights are available on Hugging Face under the Apache 2.0 license.
The model is built for batch speech recognition. That means it works on recorded audio rather than a live stream. It uses 4 billion active parameters and 26 billion parameters in total. Salesforce says the model is strong on English and also supports multiple languages.
How the model ranked in public tests
As of November 12, 2026, Salesforce says SOV-Transcribe-1 ranks first among open weight models on Artificial Analysis. The ranking is for batch speech to text. It ranks seventh among all models in that test.
On the Artificial Analysis test, the model scores 2.28 percent for missed or changed words. Salesforce compares that result with NVIDIA Nemotron 3.5 Transcribe at 2.81 percent. Both models have the same number of active parameters. It also compares the result with Mistral Voxtral Small at 2.77 percent. That model has 24 billion parameters.
VoiceCodeBench is a private leaderboard from BeSimple. It checks whether a system captures key terms such as phone numbers, emails, URLs, and code symbols. Salesforce says SOV-Transcribe-1 passes 76.3 percent of recordings and recovers 93.4 percent of entities. That result ranks first among open weight models and second among 23 systems overall.
Accents, noise, and real world audio
Salesforce also reports results on AppTek Call Center Dialogues. That is a long form test of agent and customer calls across 14 English accents. SOV-Transcribe-1 achieved an 8.21 percent average rate of missed or changed words. It had the strongest result on 8 of the 14 accents. Those accents include Australian, British, Scottish, Welsh, Irish, Indian, Mexican, and Singaporean English.
HALAS is a human annotated test of earnings call audio. On that test, the model scored 25.69 percent for missed or changed words. Salesforce says that result is stronger than the other open weight models it lists in the post. The list includes models from Alibaba, Mistral, NVIDIA, and IBM.
Style control and term hints
The model transcribes speech word for word by default. It can also follow plain language instructions about format. Users can ask it to write numbers as digits or to leave out fillers. Salesforce says this helps teams that need a verbatim record, such as a deposition. It also helps teams that need clean dictated text.
On a style test called Preference ASR, the model scored 5.05 percent when it received format instructions. That result placed second among the models Salesforce lists, behind IBM Granite Speech 5.0.
Users can also give the model keywords to boost, or a short description of the audio. Contextual Earnings-22 is a test built around names of people, companies, and products in earnings calls. The keyword match score rose from 71.9 to 96.8. That was when the model received the keywords spoken in each clip. With a much longer list of terms from the full call, the score was still 94.0. That was the highest among the open weight models in the post.
Why this matters for transcription
An open weight model lets developers and enterprises run speech tools on infrastructure they control. Salesforce says teams can use SOV-Transcribe-1 to build voice agents, dictation tools, and call transcription.
Strong results on names, phone numbers, and emails matter in real workflows. A single changed digit or a missed character in an email address can send work down the wrong path. The term hinting results will interest teams that transcribe earnings calls and clinical notes.
We will keep tracking open speech models as more companies release weights that transcription teams can test and deploy.