Phonely launches Alma, voice AI trained on 10M calls

Phonely's launch of Alma addresses a common issue in voice AI: models trained on text often produce unnatural pauses and responses. By training on over 10 million real phone calls, Alma aims to sound more human and reduce the need for extensive prompt engineering. Its faster response time and lower cost compared to GPT-4.1 and GPT-5.4 could make it an attractive option for businesses deploying voice agents.

The company's emphasis on continuous improvement through real traffic is notable. Alma uses feedback from conversations to adjust its behavior without waiting for the next training cycle, potentially leading to better performance the next day. This self-improvement loop may give it an edge in maintaining naturalness and reliability across long conversations, where general-purpose models often struggle.

For businesses, Alma's compatibility with existing transcribers and text-to-speech providers lowers the barrier to adoption. Its cost efficiency and speed could be significant for high-volume call centers. However, the model's performance in real-world applications and its ability to scale beyond Phonely's own traffic remain to be seen, as the company is now opening it to external developers.