LiveKit launches expressive mode to solve AI voice agents' emotional gap

LiveKit's addition of emotional prosody addresses the 'feeling skills gap' that often degrades customer satisfaction in automated support. By formalizing how LLMs instruct TTS models to use non-verbal cues, the framework lowers the barrier for engineers to deploy human-like voice agents that adapt to caller distress or excitement. In the broader streaming and B2B ecosystem, this move signals a shift from purely functional latency optimization toward high-fidelity interaction quality as a competitive differentiator. The industry's next signal will be the adoption rate of these emotionally-aware agents in high-stakes sectors like healthcare and travel, where tone directly impacts brand loyalty.

The launch follows a period of rapid scaling for the San Jose-based infrastructure provider. Per Tech in Asia, LiveKit secured a $100 million Series C funding round in January 2026 led by Index Ventures, which valued the company at $1 billion. This capital injection has supported the development of specialized real-time tools like the Gemma 4 31B inference deployment, which LiveKit reported in July 2026 achieves a 381ms time-to-first-sentence—roughly twice as fast as competing low-latency models. The company now powers voice services for major platforms, including OpenAI's ChatGPT voice features.

Market demand for emotionally nuanced AI has intensified throughout 2026. According to a June 2026 report from Brilo, production voice AI deployments grew 340% year-over-year as enterprises moved beyond pilot programs. While 62% of consumers now express comfort with AI agents for routine tasks, approximately 73% still prefer human interaction for emotionally sensitive issues. LiveKit’s focus on expressive prosody directly targets this trust threshold, attempting to bridge the gap between automated efficiency and the high-fidelity empathy typically reserved for human agents.

Technically, the 'expressive mode' addresses a fragmented landscape where TTS providers like Cartesia and Fish Audio offer competing dialects for markup. Cartesia's Sonic 3.5, for instance, focuses on sub-40ms latency (per TexttoLab, June 2026), while Fish Audio emphasizes a library of over 60 emotion tags. By acting as a cross-provider translation layer, LiveKit's framework prevents vendor lock-in for developers who require specific emotional ranges or latency profiles for different global markets or application use cases.