Former Amazon Scientist’s Boson AI Launches Speech Model to Challenge OpenAI and Meta on Voice Costs

Santa Clara-based startup Boson AI is preparing to roll out its debut speech-to-speech model, Higgs RealTime, targeting a major pain point in the conversational artificial intelligence market: the exorbitant compute cost of live voice processing.

Founded three years ago by Alex Smola, a prominent machine learning researcher and former distinguished scientist at Amazon, Boson AI claims its full-stack system operates at one-tenth the cost of competing services built by tech giants such as OpenAI. The company has raised $70 million from backers including Singapore investment firm Temasek and Chinese entrepreneur Su Hua.

The release comes as Silicon Valley shifts heavily toward native multimodal architectures. Traditional voice assistants relied on a cascaded chain—converting spoken input into text, processing it through a standard language model, and running text-to-speech synthesis back to the user. That multi-step process introduces latency delays that disrupt the cadence of human conversation. Newer speech-to-speech models process audio tokens directly, enabling full-duplex communication where users can interrupt and speak fluidly back and forth.

However, running continuous audio streams through high-performance graphics processing units creates severe economic bottlenecks for enterprises. Smola’s strategy centers on delivering a proprietary, full-stack architecture that allows corporate clients to train custom voice and video models from scratch or run them within their own private data centers.

By giving enterprise buyers direct control over model execution and data residency, Boson AI is targeting strictly regulated sectors including finance, telecommunications, healthcare, and insurance. The ability to keep proprietary data on-premises addresses core compliance hurdles that have delayed cloud-based AI deployments across sensitive industries.

The push for real-time speech interfaces has intensified across the technology industry. OpenAI recently introduced its GPT-Live audio suite, while Chief Executive Sam Altman publicly noted that voice interaction has begun displacing text input for routine queries. Meta is pursuing a similar trajectory with its Muse Spark model, aimed at seamless language switching and hardware integration across wearable devices, while Microsoft continues expanding voice tools tailored for corporate productivity software.

Beyond raw compute economics, voice AI developers face steep technical friction around latency and tone. While a one-second pause is standard in text messaging, conversational studies show that human verbal interaction breaks down if response latency exceeds 300 milliseconds. Boson AI is training its models to interpret vocal nuances, emotional inflection, and rapid conversational pacing to minimize natural hesitation.

Smola envisions customer support and sales infrastructure shifting rapidly toward voice automation, with long-term applications extending into embodied AI and robotics equipped with visual and reasoning capabilities.

The broader AI sector continues to draw heavy capital investment across enterprise software, hardware infrastructure, and biotechnology. In recent deal activity, email security developer AegisAI secured $36 million in Series A funding led by Battery Ventures alongside Accel and Foundation Capital, while AI design platform Paper raised $34 million in a Series A led by Accel and ICONIQ. Meanwhile, biotechnology firm Scribe Therapeutics raised $129 million in its Nasdaq initial public offering, selling 8.6 million shares priced at $15.