Salesforce Introduces Launa-Omni for Enterprise Voice Conversations

Salesforce AI Research has detailed Launa-Omni, a voice model built for real-time two-way conversations in enterprise settings. The company says it developed and trained the system for Salesforce and Slack experiences, where natural dialogue must coexist with accurate instruction following and visible behavior. Salesforce reports competitive evaluation results and positions the model alongside larger voice systems. Salesforce Blog report.

Architecture links speech and language

Salesforce AI Research describes Launa-Omni as an omni voice model whose core is a text model retrofitted for audio, so speech and language share a latent space. An in-house-trained text-to-speech model outside that core produces the spoken reply. Salesforce says the model is generalizable but was developed and trained for Salesforce and Slack experiences.

The design keeps a text response visible before speech is generated, which Salesforce says lets teams inspect and correct the reply as it is formed. That matters when a conversation leads to a customer record, workflow or decision. Salesforce frames this visibility as keeping a system that acts on what it heard under human direction.

Related coverage: related Conversational AI reporting.

Prior voice architectures carried tradeoffs

Salesforce separates earlier voice systems into cascaded systems and speech-to-speech models. Cascaded systems use separate speech recognition, language and speech generation models; they keep text visible but can lose vocal detail and add pauses. Speech-to-speech models can sound smoother, but Salesforce says they can raise serving cost and reduce visibility.

Launa-Omni avoids a full transcription handoff. Salesforce argues that once speech becomes text before reasoning, pace, hesitation and emphasis are lost, even though those cues carry meaning. Its approach brings speech and language together in one audio-adapted text model, then adds the separate in-house text-to-speech model to speak the reply.

Related coverage: related Conversational AI reporting.

Built for enterprise conversation requirements

Salesforce lists responsiveness, accuracy and interpretability as enterprise voice requirements. Launa-Omni uses a real-time harness that detects speech and determines when a person has finished a turn, helping the system reply promptly. The model can be customized and fine-tuned for named entities, dates, alphanumeric identifiers and industry jargon, according to the company.

The model can be paired with a frontier model when a request needs longer reasoning, search or complex task execution. Salesforce says voice remains the interaction point while deeper work is routed elsewhere. Slack users can already experience a simplified version through Slackbot Voice Dictation, according to the post.

Related coverage: related Conversational AI reporting.

Evaluation scope and stated limits

Salesforce reports competitive results on spoken instruction following, multi-turn dialogue and customer-service tasks. It cites Audio MultiChallenge, VoiceBench and τ³-Voice, saying Launa-Omni was tested against several larger models. These results should be attributed to Salesforce's account rather than treated as our independent testing.

VoiceBench is described as testing instruction following, reasoning and safety across datasets. Our recommendation is to examine the evaluation setup alongside the visible, steerable response design. Buyers should request the failure analysis, deployment terms and operational safeguards relevant to their intended application before making a procurement decision.

Related coverage: related Conversational AI reporting.

What This Means for Practitioners

The case for Launa-Omni rests on a clear architectural claim and company-run tests. A buyer should ask whether the visible text response is sufficient for audit and correction in the intended workflow, and whether fine-tuning for domain terms covers the actual customer or employee vocabulary. Those questions are conditional, not predictions.

Researchers should ask how the model performs outside Salesforce and Slack contexts, how paired frontier models affect latency and oversight, and what safety testing would be required for the specific use case. The post supports a focused enterprise voice evaluation, not a conclusion about general deployment readiness.