Salesforce AI Research has developed Launa-Omni, a voice model for real-time two-way conversations aimed at enterprise applications in Salesforce and Slack. The model adapts a text model to process audio so speech and language share a latent space; an in-house text-to-speech model then converts the written response into speech. Text responses remain visible before they are spoken, supporting inspection and correction. Salesforce highlights responsiveness through turn detection, accuracy for names, dates, and industry terms via customization, and interpretability via visible text. Company-reported results include about 59.69% on Audio MultiChallenge, 90.60% on VoiceBench, and a 72.9% macro score on τ³-Voice for customer-service scenarios. The model is described as smaller than several competing voice systems. A simplified version is already used in Slackbot Voice Dictation. Launa-Omni can be paired with a more capable model for deeper reasoning while voice remains the interface.
Salesforce AI Research has developed Launa-Omni, a voice model for real-time two-way conversations aimed at enterprise applications in Salesforce and Slack. The model adapts a text model to process audio so speech and language share a latent space; an in-house text-to-speech model then converts the written response into speech. Text responses remain visible before they are spoken, supporting inspection and correction. Salesforce highlights responsiveness through turn detection, accuracy for names, dates, and industry terms via customization, and interpretability via visible text. Company-reported results include about 59.69% on Audio MultiChallenge, 90.60% on VoiceBench, and a 72.9% macro score on τ³-Voice for customer-service scenarios. The model is described as smaller than several competing voice systems. A simplified version is already used in Slackbot Voice Dictation. Launa-Omni can be paired with a more capable model for deeper reasoning while voice remains the interface.