Artificial Analysis: Salesforce's SOV-Transcribe-1 is the top open-weight speech-to-text model at 2.3% word error

Artificial Analysis says SOV-Transcribe-1, a new open-weight speech-to-text model from Salesforce's Missionforce Sovereign AI team, scores 2.3% on its AA-WER Index. That is #1 among open-weight models and #7 of 44 models overall, proprietary ones included. The weights are on Hugging Face.

• Against other open models: the next two, Mistral's Voxtral Small and NVIDIA's Nemotron 3.5 Transcribe, both score 2.8%.
• Size: the model has 4.6B active parameters (26.6B total). Artificial Analysis notes this is about 5x fewer active parameters than Voxtral Small (24B) and close to Nemotron 3.5 Transcribe (4B).
• By dataset: it leads open-weight models on Earnings22 earnings calls (4.1% against 4.8% for Voxtral Small) and is slightly ahead on VoxPopuli parliamentary speech (1.5% against 1.8%). On AA-AgentTalk, Artificial Analysis's voice-agent conversation set, it scores 1.8%, level with Nemotron 3.5 Transcribe.
• Speed: on Salesforce's hosted endpoint it reaches a Speed Factor of 60x, about one second to transcribe a minute of audio.

The ranking and figures are Artificial Analysis's own measurements.