AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Large Language Models

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

Phonely has launched Alma, a voice AI model trained on over 10 million phone calls, claiming faster response times and lower cost than OpenAI's GPT-4.1.

AI News Feed · 2026-09-02 03:41

The company said many AI models that power voice agents have been trained on text, which can make conversations sound slightly off during phone and voice interactions. Teams may spend hours prompting and adjusting an agent to sound organic, yet the model still lacks a natural understanding of conversational flow.

Alma provides response times to the first token of under 185 milliseconds, compared with roughly 500 milliseconds for OpenAI Group PBC's GPT-4.1. The shorter pause reduces the unnatural gaps that can occur when a model looks up information on the internet, in enterprise knowledge bases, or within itself. Phonely added that Alma costs 55 cents per blended million tokens, making it 84% cheaper than GPT-4.1 at $3.50 per million tokens 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, Alma can continuously improve how voice agents perform. It can identify where conversations break down and use that as a feedback loop to teach the model how to adjust 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 a 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. Alma can work with existing transcribers and text-to-speech providers, allowing teams to adapt their current technology stack. Because the model is pretrained to sound like a person, it also reduces the need for prompt engineering.

“[Alma] is already answering millions of Phonely calls monthly, and now anyone building in voice can use it,” added Bodewes.