Research

Announcements and research on multimodal data and evaluations

• Voice Benchmarks Hear, but Don’t Listen

  We compared recorded human speech with audio from six commercial TTS systems to test whether voice agent benchmarks hear what real users sound like.

• The Naturalness Problem

  We recently partnered with Sesame on TurnBench, an open benchmark for measuring turn-taking in spoken dialogue.

• Mundo raises $24M to build the data layer for perceptual intelligence

  Today, we’re excited to announce Mundo’s $20 million Series A, led by GreatPoint Ventures, with participation from Y Combinator, Next Frontier Ventures, and E12 Ventures. Together with our previously unannounced $4 million seed round, Mundo has now raised $24 million.

• Smart Turn v3.2 Training Dataset

  A multilingual audio corpus for training native turn detection, with audio samples contributed by Mundo.