Salesforce unveils Launa-Omni voice AI model for enterprises

Salesforce AI Research has developed Launa-Omni, a real-time voice model for two-way conversations aimed at enterprise use in Salesforce and Slack. It processes audio inside the language model, preserving pace, hesitation and emphasis, and shows text responses before they are spoken. A simplified version already powers Slackbot Voice Dictation.

• Scored 59.69% on Audio MultiChallenge versus a top public score of 60.40%
• VoiceBench: 90.60%; τ³-Voice: 72.9% under a custom evaluation setup
• Generated text is visible before speech, letting teams review answers
• Smaller than many rivals and can pair with a stronger model for complex tasks