The Mason-Alberta Phonetic Segmenter: a forced alignment system based on deep neural networks and interpolation

Area of Science:

• Computational Linguistics
• Speech Processing
• Artificial Intelligence

Background:

• Forced alignment systems are crucial for segmenting speech data, enabling large-scale corpus analysis.
• Existing systems often have limitations in boundary precision and acoustic modeling approaches.

Purpose of the Study:

• To introduce the Mason-Alberta Phonetic Segmenter (MAPS), a novel neural network-based forced alignment system.
• To evaluate two key improvements: treating acoustic models as taggers and employing an interpolation technique for precise boundaries.

Main Methods:

• Developed a neural network architecture for forced alignment.
• Implemented an acoustic model as a tagger, acknowledging overlapping speech segments.
• Utilized an interpolation technique to refine boundary detection beyond the typical 10 ms limit.

Main Results:

• All MAPS configurations significantly outperformed the Montreal Forced Aligner at a 10 ms tolerance.
• Achieved up to a 28.13% relative performance increase in boundary placement accuracy.
• Observed the Montreal Forced Aligner slightly outperforming MAPS only at a 30 ms tolerance.

Conclusions:

• MAPS demonstrates superior performance in precise phonetic boundary detection compared to state-of-the-art systems.
• The study highlights potential discrepancies between acoustic model training targets and phonetic similarity, suggesting future research directions.
• Rethinking acoustic modeling and output targets may be necessary for further advancements in forced alignment.