Before generation, a normalizer rewrites numbers, dates and symbols into spoken form, so "$20" becomes "twenty dollars". The company says nearly half of TTS errors are fixed by the normalizer. Each line is then scored for accuracy (WER), naturalness, noise, and pronunciation, and flagged lines are auto-fixed or regenerated. Its blog says the product covers 100+ TTS models, and that one dictionary update propagates across every provider. That is a bigger figure than the 30+ voices named in the launch post. The team comes from Podonos, which offers a human-in-the-loop platform for analyzing voice AI models. Serac Ventures led its round, with Naver D2SF and KAIST Ventures joining. Its earlier customers include Resemble AI, Play AI, and Sanas AI. Tools such as Pictory and Lumen5 already let users fix pronunciation by hand, one word at a time. For founders, this points to a QA and routing layer forming on top of commodity TTS. Value is moving away from the voice model and toward checking the output, and switching providers costs less when the dictionary and the quality checks sit outside any single vendor.