ndebvu/Chichewa-Speech-dataset

Chichewa speech-dataset(On-going) Description This project aims to build a high-quality TTS model for Chichewa. So far, we have successfully collected and processed 4 hours of transcribed speech data, a foundational step in this initiative. However, significantly more data is required for robust, natural-sounding synthesis. We aim to develop a dataset of 50 hours as studies in speech synthesis (e.g., Tacotron 2, FastSpeech) indicate that high-quality, intelligible TTS models typically require at least 50-100 hours of clean, annotated speech. For low-resource languages, training with smaller datasets (e.g., 4-10 hours) results in poor pronunciation, unnatural prosody, and limited generalisation.The dataset will be made public as soon as the thesis under which the dataset is part of, is defended. The screenshot below shows despicts the dataset on hugging face Chichewa Speech Dataset: