14th International Congress of Phonetic Sciences (ICPhS-14)
San Francisco, CA, USA
In our experiment we used a bigram language model and a
standard speech recogniser to test if linguistic information is
related to the position of silence, articulatory noise, background
noise, laughing and breathing in spontaneous speech.
We observed that for silence and articulatory noise the acoustic modelling is more important than linguistic information represented in the bigrams of a language model. Breathing carries useful information that can be described in a language model, because including it into the language model improves test set perplexity and recognition accuracy.
This means that precisely defined noise items add some linguistic knowledge to the language model and contribute to a better performance of an automatic speech recogniser.
Bibliographic reference. Weilhammer, Karl / Schiel, Florian (1999): "Investigation of language structure by means of language models incorporating breathing and articulatory noise", In ICPhS-14, 1649-1652.