A Comparison of speech recognizers created using manually-aligned and automatically-aligned training data
John-Paul Hosom · container not stated · 2018
Not cached yet. This record was fetched live for this request and is stored after the lookup. Reopening the page should serve the cached copy.
· OpenAlex · dissertation · 2018
Abstract
It is generally acknowledged that time-alignment of phonemes is of better quality when obtained from manual segmentation as compared to automatic segmentation. However, empirical evidence for this belief is sparse. This paper describes a controlled study of two recognizers created using manually-aligned and automatically-aligned training data. Both recognizers were trained on the digits task using telephone- channel continuous speech. The manual alignments were generated by expert labelers, while the automatic alignments were obtained from our best general-purpose forced alignment system.
The two recognizers were trained and evaluated using the same set of data and parameters whenever possible. The results of the recognizer trained on manually- aligned data were 97.54% word accuracy and 90.18% sentence accuracy. The results for the recognizer trained on automatically-aligned data were 97.24% word accuracy and 88.80% sentence accuracy. This represents an 11% reduction in error at the word level and a 12% reduction in error at the sentence level. The sentence-level results are statistically significant using McNemar''s test, with p=0.002.
Open the resolved locator · 10.6083/000000239
Metadata licence: OpenAlex metadata (CC0)