Datasets:
license: cc-by-4.0
language:
- en
- fr
tags:
- audio
- speech
Dataset Description
This dataset contains 50 short speech recordings prepared for a text-to-speech (TTS) task. The recordings were produced by a single speaker in a quiet indoor environment using the same recording device throughout. The dataset includes 40 English sentences and 10 simple French sentences.
Issues Encountered and Solutions
The primary issue encountered during data preparation was audio segmentation. Manual segmentation proved excessively time-consuming, so the “timestamp annotation” feature was employed to automate the process. A secondary challenge arose from natural pauses within sentences, which rendered automatic segmentation inaccurate; interference from sounds such as swallowing or exhaling also impacted results. This issue was mitigated by re-recording the audio with deliberate one-to-two-second silences between sentences, thereby enhancing the accuracy of silence detection. Another problem concerned character encoding errors when processing French accented characters in the metadata file. This was resolved by saving the metadata.csv file using UTF-8 encoding and avoiding spreadsheet software that might alter the file's encoding.