🇺🇸 English 🇰🇷 한국어 🇯🇵 日本語
Home › AI & Machine Learning
AI & Machine Learning

Salesforce Opens Top Open-Weight Speech Model

Salesforce released SOV-Transcribe-1 on Nov. 12, an Apache 2.0 open-weight speech-to-text model for recorded audio in seven languages. Artificial Analysis ranks it the most accurate of 16 open-weights models at 2.3% AA-WER; it does not support live streaming.

Salesforce on Thursday, Nov. 12, released SOV-Transcribe-1, an open-weight speech-to-text model that companies can run on their own servers. Independent benchmarker Artificial Analysis now ranks it the most accurate open-weights model on its non-streaming leaderboard, with a 2.3% word error rate.

A studio microphone on a stand in a dimly lit room with soundproofing behind it
Speech recognition models turn recorded audio, such as calls and meetings, into text. File photo. 사진: Shixart1985 / 위키미디어 공용(CC BY 2.0)

Key takeaways

  • SOV-Transcribe-1 is free to download under the Apache 2.0 license and transcribes recorded audio of up to 30 minutes in seven languages.
  • Artificial Analysis lists it first among 16 open-weights models, at 2.3% AA-WER, ahead of Mistral's Voxtral Small and Nvidia's Nemotron 3.5 Transcribe at 2.8%.
  • It does not handle live, streaming audio, and Salesforce's own benchmark results have not yet been backed by a technical report.

What is SOV-Transcribe-1?

Salesforce's Sovereign Intelligence team calls the model a "speech-augmented language model." According to the Hugging Face model card, it pairs a Whisper large-v3 audio encoder with Google's Gemma 4 26B A4B mixture-of-experts language model.

The model card lists 26.6 billion total parameters, with 4.6 billion active per step when the audio encoder is counted. It supports English, French, German, Spanish, Italian, Portuguese and Dutch.

The "sovereign" label refers to control. Because the weights are open, a company can keep call recordings on infrastructure it runs rather than sending audio to an outside API.

How accurate is it on benchmarks?

On Artificial Analysis' leaderboard, checked Nov. 15, SOV-Transcribe-1 is the top open-weights model at 2.3% AA-WER, a share of words transcribed incorrectly. Salesforce says that places it seventh among all systems. The leaderboard covers 65 models, and the overall leader, Alibaba Cloud's Fun-Realtime-ASR-preview, scores 1.7%.

Salesforce's launch post adds its own results. On AppTek call-center dialogues spanning 14 English accents, it reports an 8.21% average error rate and the lowest error on eight accents. On VoiceCodeBench, which checks phone numbers, emails and code symbols, it says the model recovers 93.4% of entities and ranks second of 23 systems.

On HALAS, a set of noisy earnings-call audio, Salesforce reports a 25.69% error rate, below Alibaba's Qwen3-ASR 1.7B at 26.01% and Nvidia's Nemotron 3.5 Transcribe at 31.59%.

Instructions and custom vocabulary

The model transcribes word for word by default but follows plain-language instructions, such as writing numbers as digits or dropping filler words. Users can also feed it names or terms to listen for. On a test built from earnings-call clips, Salesforce says supplying the clip's keywords raised keyword accuracy, measured as F1, from 71.9 to 96.8.

Speed figures are also Salesforce's. On one Nvidia RTX PRO 6000 Blackwell GPU, it says the model processes up to 1,039 seconds of audio per second and returns a 30-second transcript in 0.22 seconds.

Context: open speech models are closing the gap

The release follows Nvidia's open Nemotron 3.5 Transcribe on Oct. 28 and ElevenLabs' paid Scribe v3 on Nov. 10, which Artificial Analysis lists at 1.9%. On that index, the best open model now trails the second-ranked system by less than half a percentage point.

Large platforms are also pushing cheaper voice stacks, including Google's new voice models that talk while they think.

What's next

Salesforce says a technical report is "coming soon" and that native support in vLLM and Hugging Face Transformers is being upstreamed. Until then, developers use a plugin and code shipped in the repository. The model card also warns that the model can occasionally loop on a repeated phrase and recommends a cleanup step.

FAQ

Can SOV-Transcribe-1 caption live calls? No. It is built for non-streaming transcription of recorded files, up to 30 minutes each in Salesforce's testing.

Is it free for commercial use? Yes. The weights are published under Apache 2.0, a permissive license that allows commercial use.

Are the accuracy claims independent? Partly. The 2.3% AA-WER figure comes from Artificial Analysis; the call-center, entity and HALAS results are Salesforce's own.

References

#Salesforce #SOV-Transcribe-1 #Speech Recognition #ASR #Speech-to-Text #Open-Weight Model #Artificial Analysis #Hugging Face
S
Seung Jung
AI Newsway · Developer

Editor responsible for AI Newsway. Writes about AI, developer tools, cloud and productivity technology.

More by this reporter ›
Copyright © AI Newsway All rights reserved. Unauthorized reproduction, redistribution, and AI training prohibited.
Share
f X K

Comments 0

Comments are currently disabled.

Be the first to comment.