Former Amazon scientist bets on cheaper AI to disrupt voice race

A former Amazon distinguished scientist thinks the fastest way into the crowded voice AI race isn't a better model; it's a cheaper one. Alex Smola, founder of Santa Clara-based Boson AI, is preparing to release the company's first speech-to-speech model, Higgs RealTime.

Smola told Fortune that voice technology has finally crossed the threshold needed to drive the next major shift in how people interact with machines, and he's positioning Boson's models as roughly ten times cheaper to run than competitors' while remaining, in his words, "very competent to use".

While competitors may opt to keep their systems closed, the “full-stack” approach of Boson allows enterprise clients to deploy their systems within their own data centres and train bespoke voice and video models right from scratch, ensuring client control of the data residency in sensitive sectors.

Smola is a strong supporter of the open-weight, customisable model despite Boson’s eventual decision to create proprietary systems for their business clients. 

The three-year-old company has raised $70 million, funded by the Chinese entrepreneur Su Hua and the venture arm of Singapore-based Temasek, and initially focused on the financial sector, telecommunication, healthcare, and insurance.

Boson enters a market that has been experiencing competition from companies like OpenAI, Meta, and Microsoft, who are competing to develop full-duplex voice systems that allow users to interrupt the AI during conversations.

Sam Altman, CEO of OpenAI, was recently quoted saying that he has started talking to ChatGPT more than texting it because of an upgrade in the voice model that "really crossed a threshold".

GPT-Live by OpenAI, the Muse Spark Model by Meta for natural and interruptible conversations, and a new upgrade by Microsoft in their voice system which revolves around workplace productivity are the result of such a competition.

According to Smola, apart from being costly, latency remains the core technical challenge, as what is acceptable in the context of a text chat seems annoying in the case of a voice conversation.