# SOV-Transcribe-1 Benchmark Scores & Performance

> SOV-Transcribe-1 has Salesforce launch evaluation results. Voice evidence is display-only and does not produce a text-model score.

Canonical page: https://benchlm.ai/models/sov-transcribe-1

Last updated: November 12, 2026

## Model Details

| Property | Value |
|----------|-------|
| Creator | Salesforce |
| Source Type | Open Weight |
| Reasoning Type | Non-Reasoning |
| Context Window | N/A |
| Official model card | [Salesforce model card](https://huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1/blob/505d8fae243924c0d8acdc622f7f0abab344eba3/README.md) |
| Overall Score | Not scored (voice evidence only) |
| Overall Rank | Unranked |

## Family & Coverage

- Family: SOV-Transcribe-1
- Variant: base
- Benchmarks covered: 0 of 818
- Coverage note: Speech and voice results appear below; no weighted text-model benchmark rows are stored.

## Speech and voice evidence

[SOV-Transcribe-1 launch evaluation](/voice-benchmarks/sov-transcribe-1-launch-evaluation)

Salesforce ran every result except the AA-WER rows, which it quotes from an independent speech-to-text board, using vLLM with greedy decoding, the DFlash draft off, a dataset-specific prompt for each set, and a repetition cleanup step adapted from the Qwen3-ASR toolkit. The prompts differ by set (verbatim with fillers for most English sets, fillers omitted for SPGISpeech, Latin script only for Monsoon English), so compare these numbers only with runs that used the same prompt and normalizer. The English public-set mean covers seven sets and leaves out LibriSpeech, so it is not the nine-set Open ASR mean other cards report. The multilingual mean covers German, French, Italian, Spanish, Portuguese and Dutch. The AA-WER figures are that board’s own measurements as Salesforce quotes them; the rank moves as the board adds systems. The card’s results table is an image; BenchLM transcribed Salesforce’s column only and does not carry the other models’ columns. These results do not enter text-model or voice-agent rankings.

| Metric | Result |
| --- | --- |
| AA-WER Index (independent non-streaming board) | 2.28% |
| Open ASR English public-set mean WER (7 sets) | 5.44% |
| Open ASR multilingual mean WER (6 languages) | 4.88% |
| VoiceCodeBench task success rate | 76.33% |
| AppTek call-center 14-accent mean WER | 8.21% |
| Active parameters | 4.6B of 26.6B |

Salesforce Hugging Face model card retrieved November 12, 2026. Provider-reported except the AA-WER Index row, which Salesforce quotes from an independent speech-to-text board. Error rates are better lower. Voice results are display-only.

- [Evaluation source](https://huggingface.co/salesforce-sovereign-intelligence/SOV-Transcribe-1)

### Independent test: [Independent speech-to-text board](/voice-benchmarks/artificial-analysis-speech-to-text)

Independent test by an outside speech-to-text board: batch transcription of English voice-agent, parliament and earnings-call audio through each provider's API, read 2026-11-12. The index weights the three sets 50/25/25; error rates are lower-is-better.

| Metric | Result |
| --- | --- |
| Place | #7 of 64 |
| Place among open weights | #1 of 16 |
| AA-WER Index | 2.28% |
| AA-AgentTalk WER | 1.79% |
| VoxPopuli WER | 1.50% |
| Earnings22 WER | 4.06% |

Weights published on Hugging Face under Apache 2.0 (model card dated November 12, 2026), in BF16 with the MoE router in FP32. It takes 16 kHz audio up to 30 minutes (tested) and runs with Transformers (remote code) or a vLLM 0.32.0 plugin shipped in the repository, which serves an OpenAI-compatible transcription endpoint; native vLLM and Transformers support is being upstreamed. An optional DFlash draft model in its own repository speeds up decoding about 2–3× for one request at a time. It is batch transcription, not a streaming service.