Phonely Ltd. today announced the launch of Alma, a large language artificial intelligence model built for voice agents, trained on over 10 million real phone conversations.
The company said that many AI models driving voice agents have been trained on text. This can make conversations sound slightly “off” to users on the phone and in voice. Teams may spend hours prompting and adjusting the agent to sound organic, but they still lack natural understanding of conversational flow.
Alma provides sub-185-millisecond response times to the first token, compared with roughly 500 milliseconds for OpenAI Group PBC’s GPT-4.1.
This reduces the pauses that interrupt conversations and make them feel uncanny – a normal human might pause briefly while thinking, but most people begin replying before they fully form an answer. This means the pause while an LLM is looking up information on the internet, or in enterprise internal knowledge or inside itself can feel unnatural.
Phonely added that its architecture also saves money, costing 55 cents per blended million tokens, making it 84% cheaper than GPT-4.1, which costs $3.50 per million and 90% cheaper than GPT-5.4 at $5.63. In a speed comparison, Alma is 63% faster on the first token than GPT-4.1 and 82% faster than GPT-5.6.
“Every voice agent on the market is running on a model built for something else,” said co-founder and Chief Executive Will Bodewes. “We had the phone calls, so we built the model for the calls.”
The company said its advantage extends beyond individual phone calls. By building and operating the model across millions of conversations, it can continuously improve how voice agents perform.
Alma can identify where conversations break down and uses that as a feedback loop to teach the model how to adjust its responses next time. The frontier model uses real traffic to change how it interacts with a customer’s calls without waiting for the next training cycle, delivering results that help it work better the next day instead of next month.
According to Phonely, this self-improvement makes Alma the superior choice when general-purpose models struggle not just to sound natural but to keep up with voice conversations. Voice agents need to follow instructions well, respond quickly and maintain reliability across long conversations while still sounding natural at every turn.
The company said Alma can work with existing transcribers and text-to-speech providers, so teams can adapt their existing technology stack. Also, because the model is pre-trained to sound like a person, it reduces the need to prompt-engineer it to improve its behavior.
“[Alma] is already answering millions of Phonely calls monthly, and now anyone building in voice can use it,” added Bodewes.
