Salesforce's SOV-Transcribe-1: keeping call audio on your servers

By DataForge team17 November 20265 min read

Salesforce has released an open-weight speech model NZ firms can run in-house. Here's what it does well, where it falls short and how to test it.

What Salesforce released

On 12 November 2026 Salesforce published SOV-Transcribe-1, a speech recognition model with downloadable weights under the Apache 2.0 licence. The model comes from Salesforce's Sovereign AI group. Crypto Briefing reports that the group sits under Missionforce, the company's government and national security unit. Salesforce's launch post says the model lets organisations build voice agents, dictation tools and call transcription while "keeping critical intelligence on infrastructure they control".

According to the model card, SOV-Transcribe-1 joins a Whisper large-v3 audio encoder to Google's Gemma 4 26B A4B mixture-of-experts language model. The combined model has 26.6 billion parameters, of which 4.6 billion are active for any given token. It accepts up to 30 minutes of 16 kHz audio and can follow plain-language instructions about formatting, such as writing numbers as digits or removing "ums". It also accepts a list of names and technical terms to watch for. It transcribes recordings after the fact and doesn't do live, streaming transcription.

How good is it?

Salesforce reports a word error rate of 2.28% on the Artificial Analysis non-streaming benchmark. On that figure the model ranks first among open-weight models and seventh overall, ahead of Mistral's Voxtral Small (2.77%) and NVIDIA's Nemotron 3.5 Transcribe (2.81%) (Hugging Face blog). Crypto Briefing puts the seventh place among 44 evaluated models. On AppTek's call-centre dialogue set, which covers 14 English accents, the model averaged 8.21% and had the lowest error rate on eight accents, Australian English among them.

Treat those numbers as vendor claims for now. Some of the leaderboards Salesforce cites are private or run by companies with an interest in the results, and the technical report is still marked as coming soon. New Zealand English isn't among the accents Salesforce names, and te reo Māori isn't among the supported languages. The model card lists English, French, German, Spanish, Italian, Portuguese and Dutch, and says English is the strongest.

Salesforce also names a known defect. Now and then the model repeats a character or phrase until it reaches its token limit, and the card recommends a clean-up step that checks output for truncation and repetition (model card). Anyone feeding transcripts into case notes or customer records needs that check before the text is saved.

Why it matters for data sovereignty

Audio is some of the most sensitive data a business holds. A recorded sales call, a disciplinary meeting or a clinic consult can contain health details, bank numbers and opinions about named people, all in one file. At most NZ firms, transcription runs through a meeting assistant or a contact-centre add-on hosted offshore. Every recording leaves the building and sits with a third party, under that party's retention settings.

The Privacy Commissioner's position hasn't changed because the tooling has. Bell Gully's May 2026 summary of the AI guidance notes that an organisation stays accountable for personal information when it uses AI tools. It also notes that many tools keep prompts, outputs and logs by default, sometimes overseas, and that the Commissioner strongly encourages a privacy impact assessment before AI is used on personal information. A transcription vendor's retention policy and its terms on model training are therefore your problem, and they need reading line by line.

A model you can download changes the options. Run on a server in your own rack or in an NZ data centre, the recording, the transcript and the logs all stay under your control. No vendor holds a copy, and no overseas policy change can alter the deal after you sign. Open-weight transcription has existed for years through Whisper. What's new is a model that, on Salesforce's figures, gets close to the best commercial services, comes from a large enterprise software company and carries a permissive licence.

What it costs to run

The model card reports testing on one NVIDIA RTX PRO 6000 Blackwell card with 96 GB of memory. That's a workstation-class GPU you can buy outright, well short of a data centre rack (model card). Software support is still settling. The card says native support in vLLM and Hugging Face Transformers is being upstreamed. Until then, vLLM needs a plugin pinned to version 0.32.0 and Transformers needs the repository's custom code. vLLM's transcription endpoint caps each request at 600 seconds of audio by default, so the configuration needs changing before it'll accept a full 30-minute meeting.

For a firm transcribing a few hundred hours a month, one server of that class is a realistic setup to price against per-minute cloud fees. The trade-off is staff time. Somebody has to patch the server, watch the queue, update model versions and handle the repetition defect. Firms without that capacity can have a local partner host the model in an NZ facility, which keeps the data onshore without adding work for their own team. We cover the hardware side of that decision on our local AI server page.

Where it fits and where it doesn't

The model looks strongest for recorded material that's high in volume and sensitive: contact-centre calls, internal meetings, interviews, inspection notes dictated in the field, and earnings or board recordings. Salesforce's keyword feature is useful in trades and professional services, where product codes, client names and legislation references are easy to get wrong.

It's a weak fit anywhere te reo Māori appears in speech, for live captioning, and for any task that needs speaker labels, because neither the blog nor the model card mentions diarisation. Teams in those situations will need a second tool or a human in the loop. A Māori place name or a karakia at the start of a hui will probably come out mangled in a model trained without te reo, so test with recordings that include them.

Practical next steps

  • List every place your organisation records or transcribes speech, including meeting assistants that staff have switched on themselves, and note where each provider stores audio and transcripts.
  • Read the retention and model-training clauses for each of those services, and record the results in a privacy impact assessment.
  • Pick 20 to 30 real recordings that include NZ accents, background noise, te reo Māori and your own jargon, and compare SOV-Transcribe-1 with your current service on the same files.
  • Add an automatic check for repeated or cut-off output before transcripts reach any system of record.
  • Price a single-GPU server, or a hosted NZ instance, against your current per-minute or per-seat transcription spend over three years.
  • Decide who owns the patching, monitoring and model updates before anything goes into production.

If you'd like help setting up the comparison test or the privacy assessment, our AI consultancy team works with NZ and Australian organisations on exactly that kind of evaluation.

Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.

Get my free assessment How we help
Get new articles by RSS