Loka delivers natural voice agent with Amazon Nova 2 Sonic, beats GPT and Gemini

Loka has developed a natural, low-latency voice agent using Amazon Nova 2 Sonic, addressing the common frustration of robotic, slow voice assistants. Traditional pipelines rely on three separate steps — speech-to-text, LLM reasoning, and text-to-speech — which introduces compounding delays of 3 to 5 seconds. Nova 2 Sonic processes audio end-to-end, preserving tone, emotion, and urgency that text-based methods lose. In benchmarks on Big Bench Audio, it scored 87.0 for speech reasoning, surpassing GPT Realtime (83.0) and Gemini 2.5 Flash Native Audio (71.0). Its time-to-first-audio is 1.39 seconds, enabling natural barge-in behavior where users can interrupt seamlessly. Cost efficiency is also a highlight at $0.27 per hour of input audio, making it scalable for thousands of locations. Loka validated quality using an automated LLM-as-a-judge pipeline across five dimensions: response appropriateness, intent understanding, and more, confirming the agent’s ability to handle complex, multi-part requests in real-world dealership scenarios.

The approach solves both technical and economic bottlenecks. By eliminating intermediate text conversion, Nova 2 Sonic captures subtle cues like hesitation and urgency, which are critical in sales conversations. The 1.39-second response time feels instantaneous compared to the 3-5 second pauses of traditional systems, reducing customer hang-ups and protecting brand reputation. At roughly $0.27 per hour, the cost is lower than comparable real-time models, enabling businesses to deploy voice AI widely without prohibitive expenses. Loka’s architecture demonstrates that native speech-to-speech models can deliver both intelligence and speed, making them viable for high-stakes customer interactions such as automotive sales, support, and beyond.