Subscribe
0%
Illustration for: Salesforce's SOV-Transcribe-1 tops open-weight speech recognition at 2.28% WER under Apache 2.0. What it needs and where it falls short

Salesforce's SOV-Transcribe-1 tops open-weight speech recognition at 2.28% WER under Apache 2.0. What it needs and where it falls short

Salesforce released SOV-Transcribe-1, a 26.6B-parameter mixture-of-experts transcription model built on Whisper large-v3 and Gemma 4, ranked first among open-weight models on Artificial Analysis. It's free to use commercially, but it needs a big GPU, handles seven languages and its technical report isn't out yet.

Listen to this article0:00 / 4:42

Salesforce released SOV-Transcribe-1 on November 12, an open-weight speech-to-text model under the Apache 2.0 licence. Salesforce says it scores a 2.28% word error rate (WER) on Artificial Analysis's non-streaming benchmark, which it says makes it the best open-weight model there and seventh overall, proprietary services included.

Why it matters: teams that transcribe calls, interviews or meetings can now run a near-frontier model on their own hardware, commercially, with no per-minute fees and no audio leaving the building. The catch is that it's large for a speech model, it's batch-only, and nearly every number so far comes from Salesforce.

What it is

According to the model card, SOV-Transcribe-1 combines a Whisper large-v3 audio encoder, an audio projector and a decoder fine-tuned from Google's Gemma 4 26B A4B mixture-of-experts model.

Spec Detail
Parameters 26.6B total, 4.6B active incl. encoder (model card); the blog says 26B/4B
Input 16 kHz audio, up to 30 minutes per recording (tested)
Languages English, French, German, Spanish, Italian, Portuguese, Dutch
Mode Non-streaming (batch) only, not live captioning
Precision BF16; about 53 GB checkpoint (OrcaRouter)
Licence Apache 2.0

Beyond plain transcription, Salesforce highlights style steering (verbatim by default, or cleaned-up text from a natural-language instruction such as "remove filler words") and context biasing, where you pass keywords or a short description of the audio to get names and jargon right without fine-tuning.

The numbers (all vendor-reported)

Benchmark SOV-Transcribe-1 Comparison
Artificial Analysis non-streaming WER 2.28% Voxtral Small 2.77%, Nemotron 3.5 Transcribe 2.81%
VoiceCodeBench (phone numbers, emails, code) 76.3% of recordings fully correct, 93.4% of entities #2 of 23 systems, #1 open
AppTek call-center, 14 English accents 8.21% avg WER Lowest on 8 of 14 accents
HALAS (earnings-call audio) 25.69% WER Qwen3-ASR 1.7B 26.01% (OrcaRouter)
Preference-ASR (formatting with instructions) 5.05% IBM Granite Speech 5.0 2B leads at 4.70%
Contextual Earnings-22 keyword F1 71.9 → 96.8 with the clip's keywords 94.0 with a full keyword list

OrcaRouter breaks the 2.28% composite into 1.79% (AgentTalk), 1.50% (VoxPopuli) and 4.06% (Earnings22), and puts Whisper Large v3 Turbo at 4.62% on the same board. When we checked Artificial Analysis's Salesforce page on November 14, it linked to the model but its results table didn't yet list an SOV-Transcribe-1 row, so we couldn't confirm the ranking there ourselves.

What it takes to run

  • GPU: The model card gives no minimum VRAM. All of Salesforce's benchmarks ran on one 96 GB NVIDIA RTX PRO 6000 Blackwell. OrcaRouter notes the BF16 weights won't fit on a single 48 GB card.
  • Software: The recommended route is vLLM through a plugin in the repo that pins vLLM 0.32.0; native support is still being upstreamed. Transformers works with versions 5.19 up to (not including) 5.23 and trust_remote_code=True. The no-remote-code fallback hard-codes CUDA.
  • Settings: Greedy decoding (temperature 0), thinking off, instructions in the user message.
  • Long files: vLLM's server caps audio at 600 seconds by default; set VLLM_MAX_AUDIO_DECODE_DURATION_S=1800 for 30-minute files.
  • Speed: A separate DFlash draft model gives roughly 2–3x faster decoding for single requests, per the card. With it loaded, set max_num_seqs=64, because vLLM's default of 1,024 runs out of memory. Salesforce claims up to 1,039 seconds of audio per second on the server GPU.

Limits to know

  • Repetition loops: The card warns the model sometimes repeats a character or phrase until it hits the token limit, and recommends cleanup and truncation checks.
  • Missing features: OrcaRouter notes speaker diarization and word-level timestamps aren't documented.
  • Unverifiable so far: The technical report is marked "coming soon," and VoiceCodeBench is private, so neither can yet be reproduced. Benchmarks used dataset-specific prompts, the card says.
  • Hard audio is still hard: A 25.69% WER on HALAS means roughly one word in four is wrong on that set, even though it's the best open result listed.

Who should care

Companies already self-hosting Whisper for English or Western European audio, especially call centers, legal teams and researchers who can't send recordings to a cloud API, should test this against their own files. CASRAI pitches it for interview transcripts that IRB rules keep off third-party clouds. If you need live captions, other languages or a small GPU footprint, stick with what you have. And wait for the technical report or independent results before treating the 2.28% as settled.

Sources

  1. Hugging Face blog (Salesforce Sovereign Intelligence) — Introducing Salesforce SOV-Transcribe-1 huggingface.co
  2. Hugging Face — SOV-Transcribe-1 model card huggingface.co
  3. OrcaRouter — SOV-Transcribe-1: the best open speech model, hardest to run orcarouter.ai
  4. CASRAI — Salesforce Releases SOV-Transcribe-1 ASR Model casrai.org
  5. Superpower Daily — Salesforce releases open-weight transcription AI, claiming a benchmark lead superpowerdaily.com

#Salesforce#SOV-Transcribe-1#speech recognition#ASR#open weights#Hugging Face#Whisper#Gemma