Detection of phonological features in continuous speech using neural networks

Abstract

We report work on the first component of a two stage speech recognition architecture based on phonological features rather than phones. The paper reports experiments on three phonological feature systems: 1) the Sound Pattern of English (SPE) system which uses binary features, 2) a multi valued (MV) feature system which uses traditional phonetic categories such as manner, place etc, and 3) Government Phonology (GP) which uses a set of structured primes. All experiments used recurrent neural networks to perform feature detection. In these networks the input layer is a standard framewise cepstral representation, and the output layer represents the values of the features. The system effectively produces a representation of the most likely phonological features for each input frame.All experiments were carried out on the TIMIT speaker independent database. The networks performed well in all cases, with the average accuracy for a single feature ranging from 86% and 93%. We describe these experiments in detail, and discuss the justification and potential advantages of using phonological features rather than phones for the basis of speech recognition.

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

“…In one line of research, exemplified by Frankel et al (2007b), King and Taylor (2000), Manjunath and Sreenivasa Rao (2016) and Scharenborg et al (2007), attempts are made to cover a full set of features with a single multi-value classifier (with seven classes). In the second line, exemplified by Niyogi et al (1999), Pruthi and Espy-Wilson (2007), and Schutte and Glass (2005), 1 3 research concentrates on finding an optimal set of acoustic parameters for building a detector for one specific manner feature for, e.g., vowel nasalization (e.g., Pruthi and EspyWilson 2007), nasal manner (e.g., Chen 2000; Pruthi and Espy-Wilson 2004), or stops (e.g., Abdelatti Ali et al 2001;Niyogi et al 1999).…”

“…In the last decades, AFs have received increasing interest in the field of speech technology (e.g., Bitar and Espy-Wilson 1996;Frankel et al 2007b;Hasegawa-Johnson et al 2005;Juneja 2004;Juneja and Espy-Wilson 2008;King and Taylor 2000;King et al 2007;Kirchhoff et al 2000;Manjunath and Sreenivasa Rao 2016;Naess et al 2011). Instead of building acoustic models for phones, separate classifiers are trained for articulatory-acoustic features such as manner of articulation, place of articulation, and voicing.…”

“…Research on AF classification has focused on finding the ideal set of acoustic parameters for building multi-value or binary classifiers (e.g., Salomon et al 2004;King and Taylor 2000;Niyogi et al 1999;Scharenborg et al 2007) or the ideal statistical classification method (e.g., a comparison of ANNs with SVMs by Chang et al 2005). For an overview of different methods used for AF classification see King et al (2007).…”

Towards capturing fine phonetic variation in speech using articulatory features

“…An estimate of the degree of the asynchrony in feature changes in speech is given in Wester et al (2004) King and Taylor (2000). They showed that if the feature is allowed to change within a range of -/+ 2 frames from the phone boundary, the measure "all frames correct" increases significantly by 9% absolute to 63%.…”

“…For instance, artificial neural networks (ANNs) have shown high accuracies for classifying AFs (King & Taylor, 2000;Kirchhoff, 1999;Wester, 2003). Frankel et al (2004) provide a short overview of other modelling schemes, such as hidden Markov models (Kirchhoff, 1999), linear dynamic models (Frankel, 2003) and dynamic Bayesian networks (Livescu et al, 2003).…”