A Tunable Forced Alignment System Based on Deep Learning: Applications to Child Speech

Phonetic forced alignment has a multitude of applications in automated analysis of speech, particularly in studying nonstandard speech such as children's speech. Manual alignment is tedious but serves as the gold standard for clinical-grade alignment. Current tools do not support direct training on manual alignments. Thus, a trainable speaker adaptive phonetic forced alignment system, Wav2TextGrid, was developed for children's speech. The source code for the method is publicly available along with a graphical user interface at https://github.com/pkadambi/Wav2TextGrid.

We propose a trainable, speaker-adaptive, neural forced aligner developed using a corpus of 42 neurotypical children from 3 to 6 years of age. Evaluation on both child speech and on the TIMIT corpus was performed to demonstrate aligner performance across age and dialectal variations.

The trainable alignment tool markedly improved accuracy over baseline for several alignment quality metrics, for all phoneme categories. Accuracy for plosives and affricates in children's speech improved more than 40% over baseline. Performance matched existing methods using approximately 13 min of labeled data, while approximately 45-60 min of labeled alignments yielded significant improvement.

The Wav2TextGrid tool allows alternate alignment workflows where the forced alignments, via training, are directly tailored to match clinical-grade, manually provided alignments.

https://doi.org/10.23641/asha.28593971.