Salesforce is giving away a speech-to-text AI that tops free rivals on an independent test. Paid services still do better
Salesforce has released a speech-recognition AI model that any business can download, run on its own servers and use commercially without paying a license fee. The model, SOV-Transcribe-1, went up on Hugging Face, the main public hub for AI models, on November 12.
It uses Apache 2.0, a standard open-source license. That means no revenue caps, no regional bans and no need to ask Salesforce for permission.
Transcription is a large, recurring cost for call centers, sales teams, banks recording earnings calls and any company building voice assistants. Paid services that match or beat it on the independent Artificial Analysis leaderboard charge roughly $1.67 to $6 per 1,000 minutes of audio. A free model that competes on accuracy puts pressure on those prices, and on startups whose whole business is selling transcription.
It also gives buyers something many regulated firms want: keeping recordings in-house. Salesforce frames the model as "sovereign", meaning voice systems that organizations control on infrastructure they run themselves. The team that built it is listed under Missionforce, the Salesforce unit that serves US government and national-security customers.
What independent testing shows
The headline claim holds up, with limits. Artificial Analysis, a firm that benchmarks AI models independently, gives SOV-Transcribe-1 a word error rate of about 2.3% on its test of pre-recorded audio. Word error rate is the share of words a system gets wrong.
Artificial Analysis calls it the most accurate open-weight model it tests. Open-weight means the model's files are published for anyone to run. The nearest free rivals, Mistral's Voxtral Small and Nvidia's Nemotron 3.5 Transcribe, scored about 2.8% in the comparison Salesforce published.
It is not the best overall, though. Several paid services rank ahead on the same leaderboard, including ElevenLabs' Scribe v3 at 1.9% and Microsoft's MAI-Transcribe-2 at 2.0%. Salesforce said the model placed seventh overall at launch. A review of the board a day later by OrcaRouter, which sells access to many companies' AI models, put it eighth as newer paid systems were added. The board still showed seven services ahead of it on November 15.
Most of Salesforce's other claims are its own. Its strong results on call-center audio across 14 English accents come from tests Salesforce ran or chose. So do its results on capturing key details from calls and on earnings calls. One test is a private leaderboard. A technical report on how the model was trained is listed as "coming soon".
The model does not win everywhere even on Salesforce's own charts. On a test of following formatting instructions, it trailed a far smaller IBM speech model.
The catches for buyers
It handles recordings only, not live speech. That rules it out for real-time captions or voice agents that answer mid-conversation. Google and OpenAI both offer live transcription, according to a review by OrcaRouter, which sells access to many companies' AI models but does not host this one.
It covers seven European languages: English, French, German, Spanish, Italian, Portuguese and Dutch. Nvidia's free rival handles about 40, and Google's paid service more than 85, per the same review. The model's documentation also does not describe telling speakers apart, a basic need for meeting and call transcripts.
It is heavy. The files total about 53 gigabytes. Salesforce's speed tests used one Nvidia professional graphics card with 96 GB of memory. Running it today also needs add-on code, because mainstream AI serving software does not yet support it directly. No hosting provider on Hugging Face offers it yet, so there is no pay-per-minute option.
The economics favor heavy users. Salesforce reports that one such card can process more than 1,000 seconds of audio every second. If that holds in practice, OrcaRouter estimates, a firm transcribing very large volumes could beat per-minute fees. A team transcribing a few hours a week would not.
Built from Google's and OpenAI's free parts
Salesforce did not start from scratch. The model card shows it pairs the audio-listening component of OpenAI's freely released Whisper model with Google's open Gemma 4 language model. Salesforce's work was combining and retraining those pieces for business audio.
For Salesforce, which sells AI agents to companies and governments, giving the model away costs little in license revenue. It may help win customers who want sensitive audio kept off outside clouds. The company has not tied the release to any product or price.
Early uptake is modest. The model page showed about 350 downloads shortly after release. Whether it becomes a standard choice will depend on outside testers confirming the numbers Salesforce has so far reported about itself.