SOV-Transcribe-1 Hits 60x Transcription Speed on Hosted Endpoint

On 2026-11-13, the independent evaluation firm Artificial Analysis published results on X that placed SOV-Transcribe-1 at the top of the open weights speech-to-text rankings. The model, released by Salesforce and the Missionforce Sovereign AI team, posted a 2.3% word error rate on Artificial Analysis's AA-WER Index with 4.6B active parameters, ranking 7th overall among 44 models evaluated, including proprietary ones.

Artificial Analysis's breakdown shows SOV-Transcribe-1's word error rate is 0.5 points lower than the next open weights contenders: Mistral AI's Voxtral Small and NVIDIA's Nemotron 3.5 Transcribe both record 2.8%. SOV-Transcribe-1 runs on 4.6B active parameters out of a 26.6B total, and achieves higher accuracy than Voxtral Small (24B), which has roughly five times as many active parameters, as well as higher accuracy than Nemotron 3.5 Transcribe at a comparable active parameter count. Dataset-level results show SOV-Transcribe-1 posting a 4.1% error rate on Earnings22 earnings calls, against 4.8% for Voxtral Small, and 1.5% on VoxPopuli parliamentary speech, versus 1.8%. On AA-AgentTalk, a voice agent conversation dataset, it scores 1.8%, level with Nemotron. For throughput, Artificial Analysis measured a Speed Factor of 60x on Salesforce's hosted endpoint, transcribing 1 minute of audio in about 1 second. As of now, the picture rests on the evaluation data Artificial Analysis released on 2026-11-13, with no further developments reported.