Announcing Alma: A frontier LLM built for Voice

Phonely, the leading AI-powered voice agent platform, is taking on the models built by AI’s biggest players with Alma, an LLM built specifically for voice agents to drive conversations forward. Trained on over 10 million real phone conversations instead of text, Alma is designed for the messy reality of voice, like interruptions, background voices and transcription errors, rather than only clean turn-taking. It enables self-improving agents that learn from live call data over time without depending on a frontier lab’s next model. Alma already powers 100% of Phonely’s agent conversations, handling millions of calls a month for hundreds of businesses, and is now available to teams building voice agents. “Every voice agent on the market is running on a model built for something else. We had the phone calls, so we built the model for the calls. It is already answering millions of Phonely calls monthly, and now anyone building in voice can use it.” Share When customers talk to a voice agent, they can usually tell. LLMs trained for text are the dominant models driving the conversation. So although teams spend countless hours prompting and re-prompting an agent to sound natural, it can still stumble on human conversation staples, like a caller interrupting or changing their mind mid-sentence. The voice AI model problem persists. For the last few years, frontier labs prioritized making models better at coding, reasoning and long-running tasks. Those same improvements haven’t made their way to voice agents, with many still running on a general-purpose model not designed for the demands of a phone call. An agent-led phone call needs something different. Voice agents need to follow instructions exactly, respond instantly, maintain reliability across long conversations and sound natural on every turn. Models built only for text can struggle to deliver all four. That matters when approximately $2.7 billion in revenue is transacted over the phone every day across industries, including insurance, legal intake, appointment booking and sales. Every lag, missed rebuttal or unnatural interaction can put revenue at risk. Alma addresses this need by working with any transcriber and any text-to-speech provider, so teams adopt it inside their existing stack. Because it’s pre-trained to sound like a person, Alma removes most of the prompt engineering teams do today. Even when traffic spikes, its automatic failovers, multi-region routing and priority inference keep response times consistent. “Every voice agent on the market is running on a model built for something else. We had the phone calls, so we built the model for the calls. It is already answering millions of Phonely calls monthly, and now anyone building in voice can use it,” said Will Bodewes, CEO and co-founder of Phonely.