Abstract
In this work we present a Language Model (LM) that accounts for the effects of speaker workload by drawing on recent findings in cognitive psychology research. Under workload, speakers tend to shorten their utterances, but still aim to convey their message; hence they use more informative words. Inspired by the Perception and Action Cycle Method (PACM), the LM is used as a baseline dictionary that is constrained to have higher entropy. We show that the resulting LM has a power law relation to the baseline dictionary; i.e., there is a linear relation between the word log-probability under workload and its baseline log-probability. We then test for the existence of this relation in transcriptions of audio text messages (SMS) dictated while driving under different workload conditions. Significance tests were conducted using Monte Carlo simulations, with the data modeled by principal component analysis (PCA) and linear regression (LR). Based on this power law, we suggest a simple algorithm for LM adaptation under workload. Experiments show encouraging results in perplexity improvement of the LM under workload, thus providing empirical support for our model.
Original language | English |
---|---|
Pages (from-to) | 3394-3398 |
Number of pages | 5 |
Journal | Proceedings of the Annual Conference of the International Speech Communication Association, INTERSPEECH |
Volume | 2015-January |
State | Published - 1 Jan 2015 |
Externally published | Yes |
Event | 16th Annual Conference of the International Speech Communication Association, INTERSPEECH 2015 - Dresden, Germany Duration: 6 Sep 2015 → 10 Sep 2015 |
Keywords
- Language model
- Speaker workload
ASJC Scopus subject areas
- Language and Linguistics
- Human-Computer Interaction
- Signal Processing
- Software
- Modeling and Simulation