DEV Community

Cover image for Open Weights Speech Model: SOV-Transcribe-1 vs Voxtral
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

Open Weights Speech Model: SOV-Transcribe-1 vs Voxtral

On 12 November 2026 Salesforce released SOV-Transcribe-1, and it is now the most accurate open weights speech model for recorded audio. Artificial Analysis scored it at 2.3% on its AA-WER Index, ranking it 7th of 44 models including proprietary ones, ahead of Mistral's Voxtral Small and NVIDIA's Nemotron 3.5 Transcribe (both 2.8%). The verdict: pick SOV-Transcribe-1 if you can host an 80 GB-class GPU and need accurate English transcripts; pick a hosted API or a smaller model if you cannot.

TL;DR

  • SOV-Transcribe-1 is Salesforce's first open-weight speech-augmented language model, released under Apache 2.0 on 12 November 2026 (Hugging Face).
  • It is the top open-weight model on the AA-WER Index at 2.3%, and 7th of 44 overall (Artificial Analysis).
  • It has 26.6B total parameters but only 4.6B active (model card), so memory, not compute, is the constraint.
  • The BF16 weights alone come to about 53 GB (derived below), which rules out 12 GB, 16 GB and 24 GB cards.
  • It is non-streaming and covers seven languages, so it suits recorded meetings and calls rather than live captions.
  • Last verified: 13 November 2026.

Which open weights speech model is most accurate right now?

On the independent benchmark, SOV-Transcribe-1 is. Salesforce's own write-up reports a 2.28% word error rate, against 2.81% for NVIDIA Nemotron 3.5 Transcribe and 2.77% for Mistral Voxtral Small (Salesforce on Hugging Face). Artificial Analysis rounds the two rivals to 2.8% each.

Model AA-WER (Salesforce figures) Active parameters
SOV-Transcribe-1 2.28% 4.6B
Mistral Voxtral Small 2.77% 24B
NVIDIA Nemotron 3.5 Transcribe 2.81% Same as SOV-Transcribe-1

Source: Salesforce blog. The blog describes Voxtral Small as 4B against 24B active, while the model card rounds SOV-Transcribe-1 to 4.6B active.

The gap over Nemotron is about half a percentage point. That is a real lead on a benchmark of this size, but it will not look dramatic on a clean podcast. It matters more on messy audio, which is where the other results are worth reading.

How does it do on hard audio such as earnings calls and call centres?

Better than the other open models, with a caveat on absolute numbers. On HALAS, a set of real, human-annotated earnings-call recordings, it reaches 25.69% WER. Qwen3-ASR 1.7B scores 26.01%, Voxtral Small 27.35%, Nemotron 3.5 Transcribe 31.59% and IBM Granite Speech 5.0 2B 34.51% (Hugging Face blog). A quarter of words wrong is still a lot, so hard audio still needs human review.

Other published per-set results, as reported by Crypto Briefing:

  • 4.1% on Earnings22 (earnings calls)
  • 1.5% on VoxPopuli (parliamentary speech)
  • 1.8% on AA-AgentTalk (voice-agent conversations), level with Nemotron

On AppTek's call-centre dialogues, a long-form agent-and-customer set spanning 14 English accents, the model averaged 8.21% WER and had the lowest error on 8 of the 14 accents, including Australian, British, Indian and Singaporean English (BenchLM). If you serve a multi-accent English customer base, that result is more informative than the headline number.

Does it get product names and jargon right?

Largely, yes. On VoiceCodeBench, where a recording only passes if every target entity is correct, it records a 76.3% task success rate and 93.4% entity recovery. That places it first among open-weight models and second among 23 systems overall (Hugging Face blog).

Entity accuracy is the metric that tends to hurt in practice: a transcript with one wrong function name or drug name can be worse than one with several wrong filler words. The model also accepts natural-language instructions to steer transcript style and supports context biasing, so you can pass domain terms in advance (same source). Treat that as a feature to test on your own vocabulary rather than assume.

What hardware does an open weights speech model like this need?

This is where the lead comes with a cost. The model card lists 26.6B total parameters, including the audio encoder, with 4.6B active per token. The architecture pairs a Whisper large-v3 audio encoder and an audio projector with a Gemma 4 26B A4B mixture-of-experts language model. It ships in BF16, with the router in FP32.

Our arithmetic, derived from those figures rather than stated by Salesforce: 26.6 billion parameters (model card) at 2 bytes each is about 53 GB for weights alone, before the audio activations and cache. A mixture-of-experts design keeps compute low because few experts fire per token, but every expert must still sit in memory.

Practical targets are:

  1. One 80 GB card (A100 or H100).
  2. Two 48 GB cards in tensor parallel, giving 96 GB.

A 12 GB consumer card, a 24 GB card or a free Colab T4 with 16 GB cannot hold the BF16 build. As of 13 November 2026 the repository lists no official quantised or GGUF build. A community 4-bit conversion could change that, but none exists yet, so do not plan around one.

Inference runs in vLLM (the repository includes a plugin for 0.32.0) and in Transformers, where native support is being upstreamed. The authors also publish SOV-Transcribe-1-DFlash, a draft model for speculative decoding that proposes 15 tokens at a time and speeds up single-request decoding by about 2 to 3 times, with accuracy within run-to-run variation (model card).

What are the limits of SOV-Transcribe-1?

  • Non-streaming only. It is built for recorded files such as meetings, calls and podcasts. If you need live captions, look elsewhere.
  • Seven languages: English, French, German, Spanish, Italian, Portuguese and Dutch. The strongest evidence is for English.
  • Audio is 16 kHz input, with up to 30 minutes tested per file (model card). Longer recordings need chunking.
  • Rankings are a snapshot as of 12 November 2026 and will move as other labs respond.

Should you self-host it or use a hosted service?

Self-host if you have the GPU and a reason: data that cannot leave your network, steady high volume, or a need to tune the pipeline. The Apache 2.0 licence lets you do that commercially. The same logic applies to other open models we have covered, such as LongCat 2.5 under an MIT licence and the offline PLaMo 3 translation model.

If you have a 12 GB to 24 GB GPU, or only occasional transcription, a hosted tool is the more sensible choice today. Our comparison of transcription software covers those options and their pricing trade-offs. For the speech-output side of open weights, see Zonos3 Mini for browser-based voice cloning.

FAQ

Q: Is SOV-Transcribe-1 really open source?
A: The weights are published on Hugging Face under the Apache 2.0 licence, which permits commercial use. That makes it an open weights speech model; training data and full training code are not the same thing as the weights, so check the repository for what is included.

Q: Can I run SOV-Transcribe-1 on a 24 GB GPU?
A: Not in the shipped BF16 build. The weights alone come to about 53 GB by our arithmetic from the 26.6B parameter count, and no official quantised build is listed as of 13 November 2026.

Q: How much better is it than Voxtral Small and Nemotron 3.5 Transcribe?
A: On the AA-WER Index it scores 2.28% against 2.77% for Voxtral Small and 2.81% for Nemotron 3.5 Transcribe, according to Salesforce's published figures. The margin is under a percentage point.

Q: Does it support live, streaming transcription?
A: No. It is designed for non-streaming transcription of recorded audio, so it is a poor fit for live captions or real-time voice agents.

Q: Which languages does it cover?
A: English, French, German, Spanish, Italian, Portuguese and Dutch, seven in total, according to the model card.

Q: Is the benchmark ranking independent?
A: The AA-WER result comes from Artificial Analysis, an independent evaluator. Other results, such as HALAS and VoiceCodeBench, are reported by Salesforce and should be reproduced on your own audio before you commit.

Sources and corrections

Last verified: 13 November 2026. Corrections log: none yet. This article was produced with AI assistance and reviewed against the primary sources above.

Top comments (0)