Phonely launches Alma, a voice AI model trained on 10M+ phone conversations

Phonely, a technology company, has introduced Alma, a voice AI model trained on an extensive dataset of over 10 million real phone conversations. This model addresses a common issue in voice agents, where the lack of natural understanding of conversational flow can make interactions sound unnatural to users. Phonely claims that Alma's sub-185-millisecond response time to the first token is significantly faster than the roughly 500 milliseconds of its competitors, like OpenAI Group PBC's GPT-4.1.

This reduced response time helps avoid pauses that can make interactions feel stilted, as humans typically pause briefly during thought but usually begin responding before fully formulating their reply. The model's architecture also proves cost-effective, being 84% cheaper than GPT-4.1 at $3.50 per million tokens and 90% cheaper than GPT-5.4 at $5.63 per million.

Furthermore, Alma's speed surpasses that of GPT-4.1 by 63% and GPT-5.6 by 82%. Co-founder and CEO Will Bodewes explained that the company recognized the need for a model tailored to phone calls and thus developed Alma accordingly. Beyond individual calls, the continuous improvement loop from millions of conversations allows Alma to refine its performance.

The model's real-time adaptation to customer calls ensures it remains efficient over time, marking it as a superior choice for general-purpose models that struggle with both natural conversation and maintaining responsiveness in voice interactions. Phonely states that Alma can integrate seamlessly with existing transcribers and text-to-speech providers, making it easier for teams to incorporate into their current technology stack.

With Alma already handling millions of Phonely calls monthly, the company is now making the model accessible to anyone looking to build in voice technology.