Boson AI, a Santa Clara startup founded in 2023, has introduced a speech-to-speech voice model called Higgs RealTime that processes audio directly as audio, eliminating the intermediate transcription step to reduce latency and preserve vocal nuances like tone, pacing, and emotional coloring. The company argues that this fundamentally different processing architecture is necessary for production-grade voice AI, as removing the text conversion bottleneck enables new applications in real-time interactions such as customer service agents, AI companions, and live translation tools. Boson AI also released Higgs TTS 3 on June 4, 2026, its most capable text-to-speech model, supporting expressive conversational speech across more than 100 languages with zero-shot voice cloning and inline emotion control. In June 2026, the company launched the Higgs Avatar API, which takes a single still image and, paired with audio or text input, generates a real-time talking-head video. Earlier Higgs TTS 2 models were open-sourced on Hugging Face in May 2025, having been trained on over 10 million hours of audio data, and Boson AI co-hosted a Higgs Audio Hackathon with Eigen AI in Mountain View from March 20–22, 2026.