Phonetically Balanced Code-Mixed Speech Corpus for Hindi-English Automatic Speech Recognition

Phonetically Balanced Code-Mixed Speech Corpus for Hindi-English Automatic Speech Recognition Ayushi Pandey1, B M L Srivastava2, Rohit Kumar3*, B T Nellore1, K S Teja4*, S V Gangashetty1 IIIT-Hyderabad1, Microsoft Research2, NIT Patna3, MIT Manipal4 ayushi.pandey@research.iiit.ac.in, t-brsriv@microsoft.com, svg@iiit.ac.in Abstract The paper presents the development of a phonetically balanced read speech corpus of code-mixed Hindi-English. Phonetic balance in the corpus has been created by selecting sentences that contained triphones lower in frequency than a predefined threshold. The assumption with a compulsory inclusion of such rare units was that the high frequency triphones will inevitably be included. Using this metric, the Pearson’s correlation coefficient of the phonetically balanced corpus with a large code-mixed reference corpus was recorded to be 0.996. The data for corpus creation has been extracted from selected sections of Hindi newspapers.These sections contain frequent English insertions in a matrix of Hindi sentence. Statistics on the phone and triphone distribution have been presented, to graphically display the phonetic likeness between the reference corpus and the corpus sampled through our method. Keywords: code-mix speech, phonetic balance, newspaper corpus 1. Introduction Code-mixing is a frequently encountered phenomenon in day-to-day natural language communication, especially in multilingual and bilingual communities. Code-switching is considered to be the phenomenon of alternating languages at the sentential or clausal level, and code-mixing is the word-level insertions from one language into the sentential frame of another. The phenomenon is particularly preva- lent in speech communities where the native language and medium of education are recognized as two separate lan- guages. According to the census of 2001, 12.1% of the speakers in India are speakers of English as their second or third language. Additionally, the popularity of English in social media, print media, and also entertainment make English widely accessible to most such bilingual speak- ers. The ubiquitous prestige associated with English in the diglossic Indian situation also motivates Indian bilin- guals to show abundant code-mixing and code-switching patterns between English and other regional languages. The widespread usage and growth of this phenomenon of code- mixing mandates a shift in paradigm from monolingual au- tomatic speech recognition (ASR) studies into code-mixed speech recognition. However, computational studies for both textual and speech processing of code-mixing suffer from a sincere disadvantage: lack of data. In this paper, we present a Phonetically Balanced Code Mixed (PBCM) speech corpus, sampled from a standard- ized code-mixed text corpus, the Large Code Mixed (LCM) corpus. An optimal text selection procedure has been used to extract 6,126 utterances from the LCM. The PBCM corpus is currently in the process of being recorded and post-processed for speech recognition purposes at IIIT- Hyderabad. The paper is organized as follows: Section 2 describes some popular methods in corpora creation, and also mentions the development of code-mixed corpora for various language *participated in project while interning with IIIT-Hyderabad pairs. Section 3 details the procedure of optimal text selec- tion that we employed to design the PBCM corpus. Sec- tion 4 describes the recording procedure, and the progress of speech recording so far. …

Content truncated to 500 words