Phonely launches Alma, a voice AI model trained on 10M+ phone conversations
SiliconANGLE Kyt Dotson ● Covered by 2 sources
Phonely launched Alma, a voice AI model trained on 10M+ real phone calls. It’s built to sound and respond like a phone agent, not a chatbot reading script.
Based on reporting by SiliconANGLE, Kyt Dotson — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Phonely Ltd. has launched Alma, a large language model built specifically for voice agents and trained on more than 10 million real phone conversations. That training choice is the whole point. Phonely says a lot of voice systems are still powered by models trained on text, which can leave them sounding a little off when they’re on the phone and trying to hold a natural back-and-forth.
Alma is designed to answer fast, with sub-185-millisecond response times to the first token. Phonely compares that with about 500 milliseconds for OpenAI Group PBC’s GPT-4.1. The company says that gap matters because dead air is what makes these systems feel uncanny. People don’t wait around forever before replying, and a model that hesitates while it searches the web, checks internal knowledge, or reasons through an answer can sound less like a conversation and more like a system thinking out loud.
Cost is part of the pitch too. Phonely says Alma costs 55 cents per blended million tokens, which it describes as 84% cheaper than GPT-4.1 at $3.50 per million and 90% cheaper than GPT-5.4 at $5.63. In another speed comparison, the company says Alma is 63% faster on the first token than GPT-4.1 and 82% faster than GPT-5.6.
But Phonely is also betting on a feedback loop, not just a one-time model release. The company says Alma can identify where conversations fall apart and use that as training signal to improve how it responds later. Because it’s operating across millions of calls, Phonely says the system can get better the next day instead of waiting for the next training cycle. It also says Alma works with existing transcribers and text-to-speech providers, and that its voice-first training reduces the need for heavy prompt tinkering.
Co-founder and CEO Will Bodewes summed up the logic bluntly: most voice agents are running on models built for something else, and Phonely had the phone calls, so it built the model for the calls. The company says Alma is already handling millions of Phonely calls every month, and now it’s being opened up to other builders in voice.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI launch that sounds less like demo theater and more like product discipline. Training on real phone calls, chasing faster first tokens, and cutting the prompt-engineering tax is the right instinct for voice, where awkward pauses are basically the enemy. Open or closed, the market keeps rewarding models that fit the job instead of pretending one giant brain should do everything with a smile.
Read more about this at: SiliconANGLE