A developer created a spelling game using AI voice cloning and text-to-speech technology. They used VoxCPM2 and Chatterbox models via audio.cpp to generate word recordings in English, French, and Norwegian, but faced challenges with audio quality—words had incomplete beginnings or endings. They improved results by providing reference recordings and embedding isolated words within sentences to give the model context, then needed to precisely cut the audio at the right point to extract clean word recordings.