Salesforce's SOV-Transcribe-1 Beats Every Open-Weight Speech Model at 2.3% Error Rate

Salesforce’s SOV-Transcribe-1 now ranks first among open-weight speech-to-text models on the Artificial Analysis leaderboard. Its 2.3% AA-WER beats Mistral’s Voxtral Small and NVIDIA’s Nemotron 3.5 Transcribe, both at 2.8%. Across the leaderboard’s 44 open and proprietary systems, SOV ranks seventh overall.

AA-WER is Artificial Analysis’s aggregate word error metric, which counts substitutions, deletions, and insertions relative to reference transcripts. Lower scores indicate fewer errors. The Artificial Analysis benchmark covers non-streaming transcription.

4.6B active parameters, 2.3% AA-WER

SOV-Transcribe-1 contains 26.6 billion total parameters and activates 4.6 billion for each token. Its architecture combines a Whisper large-v3 audio encoder, a projector that converts audio features into a representation the language model can process, and a Gemma 4 26B A4B mixture-of-experts decoder.

The mixture-of-experts router sends each token through selected parts of the decoder. Voxtral Small activates 24 billion parameters, so SOV uses about one-fifth as many active parameters per step. End-to-end performance also depends on the audio encoder, sequence length, routing overhead, batching, and serving software.

Memory requirements follow the 26.6-billion-parameter total because the deployment must hold every expert. BF16 weights alone occupy roughly 53 GB before the runtime adds activations, cache, and other overhead.

Instructions shape the transcript

SOV accepts natural-language instructions alongside audio, allowing one model to handle transcription and several common cleanup tasks. Its supported capabilities include:

• Entity recognition: Targets names, phone numbers, and email addresses while handling accents and background noise.
• Style control: Produces verbatim or cleaned transcripts with specified punctuation, capitalization, and number formatting.
• Context biasing: Accepts names, domain vocabulary, or a topic description to improve recognition of specialized terms.
• Multilingual transcription: Supports English, French, German, Spanish, Italian, Portuguese, and Dutch, with audio inputs up to 30 minutes.

Developers steer the output with prompts similar to those used for chat models. The instruction Transcribe the audio verbatim, outputting only spoken words in sequence. requests a raw transcript. A cleanup instruction can direct the model to remove fillers and repeated phrases.

DFlash trades draft tokens for speed

Salesforce reports two performance measurements on a single RTX PRO 6000 Blackwell Server Edition GPU. A BF16 throughput test with the DFlash draft model processes up to 1,039 seconds of audio per wall-clock second. A separate latency test returns the transcript for a 30-second clip in 0.22 seconds. Salesforce’s hosted endpoint reached a 60x speed factor, equivalent to processing one minute of audio in about one second.

DFlash uses speculative decoding, proposing 15 tokens for the main model to verify in parallel. The published latency buckets show speedups ranging from about 1.8x to 2.9x, with accuracy changes remaining within normal run-to-run variation.

Published performance covers non-streaming workloads. Concurrency, batching, audio duration, prompt length, and deployment configuration will determine production throughput and tail latency.

From weights to an API

The published weights use the Apache License 2.0. Salesforce recommends vLLM for serving and includes a plugin that adds model support to vLLM 0.32.0. Once installed, the plugin exposes an OpenAI-compatible transcription endpoint.

After installing the repository’s vLLM plugin, a minimal server and request can be launched with:

vllm serve salesforce-sovereign-intelligence/SOV-Transcribe-1 \
  --max-model-len 32768

curl http://localhost:8000/v1/audio/transcriptions \
  -F model=salesforce-sovereign-intelligence/SOV-Transcribe-1 \
  -F file=@meeting.wav \
  -F temperature=0

Transformers users can load the model through AutoModelForMultimodalLM with trust_remote_code=True. That option executes Python code from the model repository, so production deployments should review the code and pin a tested revision.

The vLLM endpoint accepts up to 600 seconds of audio per request by default. Setting VLLM_MAX_AUDIO_DECODE_DURATION_S=1800 exposes the model’s 30-minute input limit.

Closed services retain a 0.6-point lead

Three proprietary systems remain ahead of SOV on the overall leaderboard:

1. Fun-Realtime-ASR-preview: 1.7% AA-WER
2. ElevenLabs Scribe v3: 1.9% AA-WER
3. Microsoft MAI-Transcribe-2: 2.0% AA-WER

SOV’s 2.3% score reduces the previous 2.8% open-weight benchmark by 0.5 percentage points, an improvement of roughly 18%. The remaining difference between the leading proprietary system and the leading open-weight model is 0.6 percentage points.

Production limits to plan around

Production tests should cover three known constraints:

• Repetition loops: Rare generations may repeat a character or phrase until reaching the token limit. Salesforce applies a post-processing step adapted from the Qwen3-ASR toolkit.
• Decoding configuration: The authors recommend greedy decoding with the model’s thinking mode disabled.
• Long recordings: Splitting audio at speech boundaries provides more reliable results than sending unsegmented 30-minute windows.

Self-hosted voice agents, meeting tools, and call-center analytics can use SOV to combine transcription, context biasing, and output formatting in one deployment. Teams still need to budget for all 26.6 billion weights, handle repetition failures, and benchmark their own languages, accents, vocabulary, concurrency, and audio conditions before replacing an existing pipeline.