#292 Voice AI Data for Lower-Resource Languages with NCSpeech

Dmitrii Sandzhiev and Iurii Agafonov, two of the Co-Founders of NCSpeech, join SlatorPod to talk about building speech AI datasets in emerging markets, scaling data collection through superapps, and addressing the quality and infrastructure challenges behind voice AI.Dmitrii explains that NCSpeech turns idle time in superapps into AI training data by rewarding drivers, riders, and passengers for completing data collection tasks. The company connects AI labs, enterprise R&D teams, and sovereign AI initiatives with real-world audio, image, and video data collected through the platform’s partners.Dmitrii sees particularly strong demand for speech datasets in emerging markets, where languages, dialects, and code-switching remain underrepresented in training data. In Malaysia, for example, speakers frequently switch between Malay, English, Mandarin, and local dialects.Iurii highlights the technical challenges behind producing training-ready datasets, including controlling recording conditions, verifying speaker consistency, detecting synthetic submissions, managing local data storage requirements, and processing large volumes of audio.Alongside collecting data on demand, NCSpeech is building reusable dataset libraries and developing its own models. Dmitrii shares how its Kazakh speech recognition model now outperforms available open solutions, demonstrating its technical capabilities and attracting potential customers.Looking ahead, Dmitrii and Iurii discuss NCSpeech plans to expand into more countries and apps, secure longer-term data customers, strengthen their US presence, and raise a seed round.