Salesforce's voice model speaks only after it writes: a smaller model aims at enterprise conversation you can inspect
Synopsis
Salesforce AI Research has built Launa-Omni, a voice model for real-time enterprise conversation that keeps speech and language in one audio-adapted text model and uses a separate in-house text-to-speech model to speak the reply. Unlike pipelines that chain speech recognition, a language model and speech generation, speech is not transcribed before reasoning, so pace, hesitation and emphasis survive the first handoff. The model is significantly smaller than several larger voice models and reports competitive results on spoken instruction following, multi-turn dialogue and customer-service tasks; a simplified version already powers Slack's voice dictation.
Interpretation
Enterprise voice agents now have an option that keeps speech and language in one model, with the reply visible as text before anything is spoken so teams can inspect and correct it. Earlier voice systems mostly followed one of two paths: separate speech recognition, language and speech generation models run in sequence, or a single model that takes speech in and produces speech out. The cascaded path keeps text visible at handoffs but can lose details from the original voice and add pauses; the speech-to-speech path sounds smoother but can cost more to serve and hide how a reply was formed. Launa-Omni reports competitive results on spoken instruction following, multi-turn dialogue and customer-service tasks against several significantly larger models.
Speech is no longer transcribed into text before the model reasons about it, so pace, hesitation and emphasis reach the reasoning stage. A cascaded system turns speech into text at the first handoff, and whatever a transcript cannot hold is discarded there. The core is a text model retrofitted for audio so speech and language share a latent space; outside it, an in-house-trained text-to-speech model produces the spoken reply. The model was developed and trained for Salesforce and Slack experiences, and a real-time harness detects speech and decides when a person's turn is complete.
Work beyond conversation can be routed to a frontier model while voice stays the point of interaction. A single voice model is not suited to longer reasoning, search or complex task execution on its own. Launa-Omni can be paired with a frontier model behind the experience, and Slack users can already try a simplified version in the Slackbot Voice Dictation speech-to-text feature.
Perspective
The report covers the model structure, evaluation categories and product placement that Salesforce AI Research published. Inspectability comes from the reply forming as text before it is spoken, which suits enterprise conversations that people need to review and correct; the version available in Slack is a simplified one used for speech-to-text dictation.
Results are described as competitive against several significantly larger models, without specific numbers. The model is called generalizable but was developed and trained for Salesforce and Slack experiences, and when paired with a frontier model the deeper work is handled by an external system.