Phonely's Alma: a voice LLM trained on real phone calls

Phonely launched Alma today, a language model built for one job: talking on the phone. It was trained on more than 10 million real phone conversations, and the pitch is that general-purpose models are the wrong tool for a channel where people interrupt, cough, and get transcribed wrong. If you run a business where the phone still rings, the numbers are worth reading closely.

What actually happened

Alma targets the two things that decide whether a voice agent feels human: latency and unit cost. Phonely reports sub-185ms to first token, against roughly 500ms for GPT-4.1 — about 63% faster — and around 200ms at the 99th percentile. On price it lists $0.55 per blended million tokens versus $3.50 for GPT-4.1, roughly 84% cheaper (SiliconANGLE, MarTech Series).

The model was built for the mess: interruptions, background voices, garbled transcription. It drops into an existing stack — you keep your own transcriber and text-to-speech provider — and it already runs 100% of Phonely's own agent conversations, millions of calls a month. It is now available to outside teams building voice agents.

Two caveats we will not paper over. Every figure here is vendor-published; there is no independent benchmark yet. And the customer result being circulated — a 3% conversion lift worth $300,000 a month — is a single case supplied by the vendor, not a pattern you should plan around.

Why voice agent latency matters for your business

The gap between a 500ms model and a 185ms model is not a spec-sheet detail. It is the pause where a caller says "hello?" and decides they are talking to a robot. On the phone, the feel of the response arrives before the content does, and that feel is set almost entirely by time-to-first-token and by how the system behaves when someone talks over it.

The cost math is the other half. At $3.50 per million tokens, a high-volume front desk quietly becomes a line item somebody questions. At $0.55, the calls you were routing to voicemail after hours become calls you answer. That is the actual decision — not "should we have AI," but "which calls are now cheap enough to answer."

What we would not do is rebuild your call flow around one model's API. A voice stack has four swappable pieces — telephony, transcription, model, TTS — and the model is the piece most likely to be beaten on price or latency within two quarters. Keep it behind an interface. Then a launch like this is a config change and an A/B test, not a migration.

Sources: SiliconANGLE, MarTech Series.