eduhk-compling/1153749_BianHaochen

Dataset Description

This dataset consists of high-quality voice recordings of 40 carefully selected sentences extracted from Zhu Ziqing’s renowned prose essay “Moonlight over the Lotus Pond” (荷塘月色). The recordings were made with clear, natural, and expressive reading style, capturing the quiet, delicate, and poetic atmosphere of the original work. The total audio duration exceeds 3 minutes, with optimized sample rate and bit depth to ensure high fidelity. This dataset is specifically designed for fine-tuning Text-to-Speech (TTS) models, but it is also highly suitable for tasks such as Chinese speech synthesis research, prosody analysis, intonation studies, acoustic feature analysis of literary language, and other related linguistic or literary applications

Issues Encountered & Solutions

During the preparation of this dataset, I encountered two main challenges: First, the initial recordings contained noticeable background noise and breath sounds. By carefully applying the noise reduction techniques outlined in the guidelines, I was able to significantly decrease the noise, greatly improving the clarity, purity, and overall listening quality of the voice. Second, due to the long, flowing, and rhythmically rich sentences typical of Zhu Ziqing’s writing style, automatic segmentation and alignment produced several errors — some transcriptions did not perfectly match the natural pauses, breathing points, and phrasing in the spoken performance. To address this, I listened to each segment again attentively and manually fine-tuned the text entries in the metadata.csv file.