A Praat script for unsupervised phoneme segmentation based on spectro-temporal representation

Unsupervised phoneme segmentation can be useful for corpus-based phonetic research and speech technology. Despite numerous previous studies, software that is straightforward and lightweight to install and execute is hard to find. In response, I present a Praat script which converts input recording to a spectro-temporal representation, measures how the representation differs between either side of each frame, and draws boundaries at peaks of the difference measure. The user can choose from various types of representation (e.g. mel spectrogram, cochleagram) and measure (e.g. cosine distance, spectral transition measure) as well as configure how they are computed and interpreted for peak detection. Experiments suggest that running the script with the log mel spectrogram and spectral transition measure identifies phonetic boundaries at reasonable performance levels: e.g. F-score = 0.771 on the standard TIMIT test set compared with F-score = 0.848 using the Montreal Forced Aligner (McAuliffe et al., 2017) and F-score = 0.837 using a state-of-the-art semi-supervised model based on deep learning (Kreuk et al., 2020). The script is available at the author's website for download as a single file and can be executed via Praat which many phoneticians should already have on their computer.