Broad phonetic classification using discriminative Bayesian networks

High-dimensional linear representations for robust speech recognition

“…We would seek to further improve the results by incorporating techniques used by other authors, in particular the use of commitee classifiers to combined a number of representations with different parameters. Additionally a hierarchical classification could be implemented to reduce broad phoneme class confusions [5,19,20]. There is also scope for further tuning within the method presented by weighting the sector sum and frame average, or allowing the number of frames to be different for each sector.…”

Rethinking classification results based on read speech, or: why improvements do not always transfer to other speaking styles

“…For building AF classifiers, different statistical classifiers have been used, such as artificial neural networks (ANNs) (e.g., King and Taylor 2000;Kirchhoff 1998;Scharenborg et al 2007;Siniscalchi et al 2008) and FFNNs (Manjunath and Sreenivasa Rao 2016), HMMs (e.g., Kirchhoff 1999;Manjunath and Sreenivasa Rao 2016), linear dynamic models (Frankel 2003), k nearest neighbor (k-NNs) (Naess et al 2011) and dynamic Bayesian Networks (e.g., Pernkopf et al 2009;Frankel et al 2007b;Jyothi 2013). Niyogi et al (1999), Pruthi and Espy-Wilson (2007), Yoon et al (2010), Scharenborg et al (2007), and Schutte and Glass (2005) used support vector machines (SVMs) for the classification of articulatory-acoustic features, among other reasons because these show good generalization from a small amount of high-dimensional training data.…”

“…This mapping is quite plausible, since the difference between a release realized with friction and a release realized as burst and friction is only salient within the first milliseconds of the release, i.e., the steepness of its amplitude rise. In order to distinguish friction from fricatives and releases from plosives that might be realized with a burst, plosives are mapped on to the sequence 'closure' and 'burst+release' (Frankel et al 2007b;Pernkopf et al 2009). We also opt for the latter sequence of values, for two reasons.…”

“…Even though it is well known that TIMIT (Garofolo 1988), a read speech corpus of American English, is non-generic, it continues to be the most popular corpus for research into AF classification American English (e.g., Pernkopf et al 2009;Pruthi and Espy-Wilson 2004;Siniscalchi et al 2007). The main reasons for using TIMIT are the high number of speakers, the quality of the manually created phonetic transcriptions and the large set of phonetic labels.…”

Algorithms to Automate Estimation of Time Codes for Captioning Digital Media

“…This should facilitate the process of matching these phonemes to their occurrence in the transcript. The approach of combining ZCR and wavelet analysis has been used to classify segments of speech signals into broad phonetic categories including silence, voiced, unvoiced, and plosive release [4]. The results of the CWT were more straightforward and easier to interpret than the spectrogram results.…”