Salesforce unveils Launa-Omni, a voice AI built for business calls

Salesforce has unveiled Launa-Omni, a voice AI model made for business conversations. It listens and talks in real time. Its standout feature is that it shows the text of its answer before it speaks, so a person can read it and fix it if needed. Salesforce AI Research, the company's research arm, built it.

What it is

Launa-Omni supports two-way spoken conversations between people and AI. Salesforce says it is aimed at business uses, including experiences inside Salesforce and Slack. The company says it is meant to bring accuracy, responsiveness and transparency to natural conversation, three things that matter for business tasks.

How it works

Many voice AI systems chain three separate tools: one turns speech into text, one works out a reply, and one speaks it. That makes the text easy to inspect, but it can add delays and lose details from the original speech. Other systems go straight from speech to speech. They sound natural, but it is harder to see why they said something, and they can cost more to run.

Launa-Omni takes a middle path. A text-based AI model was adapted to understand audio directly. A separate speech model, trained in-house, then speaks the reply. Salesforce says this keeps cues such as pace, hesitation and emphasis, which are lost when speech becomes a plain transcript.

Why businesses may care

• Faster turns: it detects when a speaker has finished, so replies come without awkward pauses.
• Better accuracy: it can be tuned for names, dates, ID codes and industry terms.
• Easier checking: teams can read the planned answer before it is spoken.

Salesforce says a misheard detail can damage customer records or workflows, which is why the accuracy point matters. It also says Launa-Omni can be paired with a more capable AI model for tasks that need deeper reasoning, searching or complex actions. Voice stays the main interface while the other system does the heavy work.

The test scores

Salesforce reported results on three public tests. On Audio MultiChallenge, it averaged 59.69 percent over three runs, just under the top public score of 60.40 percent in the company's comparison. On VoiceBench, which covers instruction following, reasoning and safety, it scored 90.60 percent. On τ³-Voice, a customer-service test across retail, airlines and telecommunications, it scored 72.9 percent using a custom setup.

What is not known yet

These figures are reported by Salesforce itself, and companies can pick the scores that flatter their models. The published results do not show that Launa-Omni beats every rival or that it avoids errors. Salesforce also says the model is much smaller than several competing voice models. A simpler version already powers voice dictation in Slackbot, according to the company.