Acoustics-Phonetic feature based Speech Recognition of Bhojpuri Dialect

Author: Manish Kumar Singh (University Varanasi)
Speaker: Manish Kumar Singh
Topic: Ethnography of communication
The (SCOPUS / ISI) SOAS GLOCAL CALA 2019 General Session


Bhojpuri (ISO639-3 BHO)[1] is a non-scheduled language widely spoken in western Bihar and eastern Uttar Pradesh besides some other parts of India[2]. Bhojpuri spoken in different parts exhibit distinct characteristics causing variations across the different varieties that influence the speech characteristics. This is one of the reasons for the difficulty that Bhojpuri faces for its standardization. Athough, these different varieties of Bhojpuri also share commonalities which make it different from Hindi and other sister languages[3]. In this paper, we attempt to formulate certain strategies, based on speech styles of speakers, to identify individual characteristics of the varieties of Bhojpuri. We have collected data from 15 Bhojpuri speakers from 5 regions (Patna, Varanasi, Bhojpur-Ara, Mirzapur and Ballia). We have tried to include all types of syllabic structures in this data. For the purpose of this study, we describe a method that uses acoustic-phonetic features to identify Bhojpuri dialects. We have used some statistical tools, e.g. PLP (Perceptual Linear Prediction coefficient) and MEL-scale filter bank with PLP (MF-PLP) to extract features from spoken utterances[]. We have further used an auto-associative neural network (AANN) for capturing long range features with the help of output of these tools[4]. We have then analyzed feature adequacy of AANN to model the intrinsic characteristics of speech features of a dialect[5]. The study will have significant consequences for several applications including documentation and resource building for technology developments.


[1] “,”

[2] R. B. Misra, Sociology of Bhojpuri language. Swasti Publications, 2003.

[3] D. Mishra and K. Bali, “A comparative phonological study of the dialects of hindi,” Proceedings of ICPhS XVII, Hong Kong, pp. 17–21, 2011.

[4] S. Sinha, A. Jain, and S. S. Agrawal, “Speech processing for hindi dialect recognition,” in Advances in Signal Processing and Intelligent Recognition Systems. Springer, 2014, pp. 161–169.

[5] B. Yegnanarayana and S. P. Kishore, “Aann: an alternative to GMM for pattern recognition,” Neural Networks, vol. 15, no. 3, pp. 459–469, 2002.

Keywords: Bhojpuri, Variety vs. Variations, Syllabic Identification, Acoustic-phonetic feature, MF-PLP, and AANN