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

Phonely has unveiled Alma, an advanced voice AI model trained on over 10 million genuine phone conversations. Unlike many existing AI models that have been trained on text data, Alma's training on actual voice interactions allows it to generate conversations that sound more natural to users. With sub-185-millisecond response times for its first token, Alma outpaces OpenAI's GPT-4.1, which takes around 500 milliseconds.

The model's efficiency translates to cost savings, being 84% cheaper than GPT-4.1 and 90% less expensive than GPT-5.4. Phonely's co-founder and CEO, Will Bodewes, highlighted that Alma was built specifically for phone calls, offering a significant advantage over general-purpose models. Beyond individual interactions, the model continuously learns from millions of conversations, identifying and addressing conversational breakdowns.

This self-improvement allows Alma to adapt its responses in real-time, ensuring better performance across long phone calls. Alma's compatibility with existing transcribers and text-to-speech services means teams can integrate it into their current technology stacks without major overhauls. With Alma already handling millions of calls for Phonely, the company aims to empower developers building voice applications, providing them with a more reliable and natural-sounding AI solution.