eduhk-compling/11537527_LiangJingYi

Dataset Description

This dataset is a clean Text-to-Speech (TTS) dataset in Mandarin Chinese, designed for training and evaluating TTS models. It consists of 45 phonetically diverse sentences (approximately 4 minutes of audio) recorded in a quiet environment using a consistent microphone setup. Each sentence is meticulously segmented into individual WAV files (16-bit PCM, 44.1kHz) and accompanied by exact textual transcriptions. The sentences cover a wide range of linguistic features, including varied tonal combinations, initials/finals, sentence structures, and contextual themes (daily life, technology, culture, nature, etc.). The dataset is structured for programmatic access with a standardized metadata.csv file linking each audio file to its corresponding text, ensuring it is ready for direct use in TTS pipelines.

Issues Encountered & Solution

Despite recording in a quiet room, subtle environmental noises (e.g., computer fan, distant traffic) were present. To address this, I did some noise reduction in Praat, selecting and removing noise at specific frequencies. However, to ensure clear sound, I didn't make too many modifications. In addition, I manually reviewed and adjusted segmentation boundaries by zooming into waveform views in Praat, ensuring each audio segment began with a slight breath or initial consonant and ended with a natural intonation drop, while strictly controlling leading/trailing silence to between 100–200 milliseconds. Although time-consuming, this approach ensured high purity and consistency at the sentence level, making the data more suitable for TTS model training.