Phonely launches Alma voice model for faster calls

Phonely has launched Alma, a large language model built for voice agents that already handles all calls on the company's platform.

Trained on more than 10 million phone conversations, Alma is designed to handle interruptions, background speech and transcription errors that can disrupt automated calls.

The San Francisco-based company is now making the model available to other teams building voice agents after using it internally across millions of calls each month for hundreds of businesses.

Voice systems used in customer service, sales, legal intake and appointment booking have often relied on general-purpose language models developed primarily for text tasks. Phonely argues that phone calls require a different approach because agents must respond quickly, follow instructions consistently and remain natural over long exchanges.

Phonely said Alma delivers a first token in under 185 milliseconds, compared with about 500 milliseconds for OpenAI's GPT-4.1. At the 99th percentile, Alma responds in about 200 milliseconds, while GPT-4.1 takes about 2,000 milliseconds, according to the company.

Cost is another part of Phonely's pitch. The company said Alma costs 55 cents per blended million tokens, compared with USD $3.50 for GPT-4.1 and USD $5.63 for GPT-5.4.

Call focus

Phonely said text-trained models can still struggle with common features of live conversations, including callers interrupting, changing direction mid-sentence or speaking in noisy environments. Alma is designed to work with existing transcription and text-to-speech systems, allowing customers to use it within their current voice technology stack.

Phonely also said the model improves through live call data. Because it runs Alma across its own customer traffic, the company can identify where conversations break down and use those signals to refine how agents behave over time.

That feedback loop happens automatically for customers already using Phonely's platform, according to the company. Organisations can also buy custom deployments with dedicated rate limits and infrastructure options.

The launch comes as businesses look for ways to automate more phone-based interactions without making calls sound stilted or delayed. Phonely said about USD $2.7 billion in revenue is transacted over the phone each day across sectors including insurance, legal intake, appointment booking and sales.

Commercial impact

Phonely linked call performance directly to business outcomes, arguing that hesitation or missed responses can affect conversions. It cited one customer that switched from OpenAI's model to Alma and saw a 3% increase in sales conversion, adding USD $300,000 in monthly revenue, according to the company.

That claim points to a broader commercial question in voice AI: whether businesses will favour specialised models over general-purpose systems as automated calling expands into revenue-generating tasks rather than basic support workflows.

Phonely said Alma now underpins 100% of its own agent conversations. The company is presenting that internal use as evidence that a model trained on voice interactions can outperform text-focused systems on speed, cost and call quality in production settings.

Will Bodewes, Chief Executive Officer and Co-Founder of Phonely, described the company's rationale for building its own model.

"Every voice agent on the market is running on a model built for something else. We had the phone calls, so we built the model for the calls. It is already answering millions of Phonely calls monthly, and now anyone building in voice can use it," said Bodewes.