Overview

Hopper builds fast inference for voice AI by training speech-to-text, text-to-speech, and speech-language models on a customer’s production calls and then serving and refining them with live traffic to keep voice agents low-latency and cost-efficient. It focuses on production workloads, optimizing time-to-first-token, efficient model serving, and continuous improvement using real call data rather than static benchmarks, with work cited on hybrid Mixture-of-Experts inference and diffusion-based TTS speeds. Hopper differentiates itself by prioritizing production data and real-time feedback through end-to-end inference and model-serving for voice agents, backed by YC and based in San Francisco. Its goal is to enable fast, accurate, and cost-effective voice AI in production by continually training on live data, deploying low-latency models, and iterating as calls flow through the system.