Iris2/11535335_RENZIMENGIris

Dataset Description

This dataset is prepared for Text-to-Speech (TTS) model training, containing 60 English sentences (covering daily conversations, common phrases, and simple declarative sentences) and corresponding segmented audio clips. The audio was recorded in a quiet language lab using a standard microphone, with a total duration of 3 minutes and 36 seconds. All audio files are in WAV format (16-bit PCM, 44100 Hz), named sequentially as s001.wav to s060. wav. A metadata.csv file is included to link each audio file with its original and normalized transcription, ensuring programmatic accessibility for TTS training.

Issues Encountered & Solution

During dataset preparation, we encountered two main problems. First, regarding audio recording, we faced two major issues. The first was unclear audio quality and excessive noise. Recording in the dormitory resulted in poor sound pickup, with noise levels exceeding requirements, affecting audio quality and making it unsuitable for TTS training. To address this, I requested the use of a studio's invisible recording equipment. Then, I repeatedly adjusted the computer's audio parameters, optimized the recording equipment settings, and adjusted the distance between the recording equipment and myself. After these adjustments, we finally recorded relatively clear audio that met the requirements. During audio segmentation, due to misreading, we forgot which sentences were repeated, leading to a discrepancy between the final number of audio files and the planned number. I manually adjusted the labels, carefully listened to each segment, moved the label boundaries to align with the correct sentences, and then re-exported the audio segments to ensure that each audio file corresponded to a complete sentence.