Misarticulated /r/ - speech corpus and automatic recognition technique

Abstract

Technique to recognize the impaired pronunciation of sound /r/ from Telugu speech signals is presented in this paper besides the speech corpus. Rhotacism is called as an inability to pronounce the sound /r/ and is one of the Speech Sound Disorders (SSD) in children. Whose SSD not diagnosed at an early stage may result ina lack of social skills. This demands an efficient automatic speech impairment detection technique, which helps the therapists to treat the patients with impairment specific procedure. Databases for the impaired articulation of /r/ in various languages are explored in this article. The shape of the envelope, timbre, Walsh Hadamard Transform (WHT), Discreet Cosine Transform (DCT) features extracted, from the Mel-Frequency Cepstral Coefficients (MFCC), to discriminate the correct and wrong articulation of /r/ are detailed. Usage of k-Nearest Neighbor (kNN), Support Vector Machine (SVM) and kohonen neural networks in various articles, for classification, are briefed. MFCC features and k-NN algorithm is used to identify the misarticulation in the Telugu language. The 80.1% classification accuracy shows that the proposed method performs good with respect to the methods detailed for other languages. Availability of acoustic databases for the impaired articulation of /r/ and subjects with such impairment restricts the performance validation of the investigated methods. This further demand the more contribution from scholars in the development of automatic techniques and databases for misarticulated /r/ in different languages.