Performance of Forced-Alignment Algorithms on Children's Speech.

Area of Science:

• Speech Science
• Computational Linguistics
• Developmental Psychology

Background:

• Acoustic measurement of speech necessitates precise signal segmentation into units like words or phones.
• Manual segmentation is labor-intensive and time-consuming.
• Forced alignment algorithms offer automated speech segmentation by aligning transcripts with audio data.

Purpose of the Study:

• To evaluate and compare the phoneme-level alignment accuracy of five distinct forced alignment algorithms on a child speech corpus.
• To provide performance benchmarks for researchers working with child speech data.

Main Methods:

• Five forced alignment algorithms were tested: Montreal Forced Aligner (with and without speaker adaptive training), Kaldi (triphone alignment), Prosodylab-Aligner, and Penn Phonetics Lab Forced Aligner.
• A corpus of speech from 42 children aged 3-6 years was used.
• Manual alignment served as the gold standard for evaluating automatic alignment accuracy and phone-onset time differences.

Main Results:

• The Montreal Forced Aligner with speaker adaptive training demonstrated superior accuracy and minimal timing discrepancies.
• Vowels exhibited the highest alignment accuracy across all tested algorithms.
• Alignment accuracy for fricative sounds improved with increasing age of the child speakers.

Conclusions:

• While the top-performing aligner approached human reliability, it did not fully achieve it for child speech.
• Forced alignment is a viable tool for child speech research, especially for vowels and fricatives (in older children), often within a semi-automated workflow requiring manual error checking.