Don't Multiply Lightly: Quantifying Problems with the Acoustic Model Assumptions in Speech Recognition

TitleDon't Multiply Lightly: Quantifying Problems with the Acoustic Model Assumptions in Speech Recognition
Publication TypeConference Paper
Year of Publication2011
AuthorsGillick, D., Gillick L., & Wegmann S.
Other Numbers3212
Abstract

We describe a series of experiments simulating datafrom the standard Hidden Markov Model (HMM) frameworkused for speech recognition. Starting with a set of test transcriptions,we begin by simulating every step of the generative process.In each subsequent experiment, we substitute a real componentfor a simulated component (real state durations rather thansimulating from the transition models, for example), and comparethe word error rates of the resulting data, thus quantifying therelative costs of each modeling assumption. A novel samplingprocess allows us to test the independence assumptions of theHMM, which appear to present far more serious problems thanthe other data/model mismatches.

Acknowledgment

This work was partially supported by funding provided to ICSI through National Science Foundation grant IIS:1015930 (“RI: Small: Exploratory Data Analysis for Speech Recognition”). Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors or originators and do not necessarily reflect the views of the National Science Foundation.

URLhttp://www.icsi.berkeley.edu/pubs/speech/dontmultiplylightly11.pdf
Bibliographic Notes

Proceedings of the Automatic Speech Recognition and Understanding Workshop (ASRU 2011), Big Island, Hawaii

Abbreviated Authors

D. Gillick, L. Gillick, and S. Wegmann

ICSI Research Group

Speech

ICSI Publication Type

Article in conference proceedings