Structural Analysis of Hindi Phonetics and A Method for Extraction of Phonetically Rich Sentences from a Very Large Hindi Text Corpus

2016 Conference of The Oriental Chapter of International Committee for Coordination and Standardization of Speech Databases and Assessment Technique (O-COCOSDA) 26-28 October 2016, Bali, Indonesia Structural Analysis of Hindi Phonetics and A Method for Extraction of Phonetically Rich Sentences from a Very Large Hindi Text Corpus Shrikant Malviya∗, Rohit Mishra† and Uma Shanker Tiwary‡ Department of Information Technology Indian Institute of Information Technology, Allahabad, India 211012 ∗Email: shrikant.iet6153@gmail.com †Email: rohit129iiita@gmail.com ‡Email: ustiwary@gmail.com Abstract—Automatic speech recognition (ASR) and Text to speech (TTS) are two prominent area of research in human computer interaction nowadays. A set of phonetically rich sentences is in a matter of importance in order to develop these two interactive modules of HCI. Essentially, the set of phonetically rich sentences has to cover all possible phone units distributed uniformly. Selecting such a set from a big corpus with maintaining phonetic characteristic based similarity is still a challenging problem. The major objective of this paper is to devise a criteria in order to select a set of sentences encompassing all phonetic aspects of a corpus with size as minimum as possible. First, this paper presents a statistical analysis of Hindi phonetics by observing the structural characteristics. Further a two stage algorithm is proposed to extract phonetically rich sentences with a high variety of triphones from the EMILLE Hindi corpus. The algorithm consists of a distance measuring criteria to select a sentence in order to improve the triphone distribution. Moreover, a special preprocessing method is proposed to score each triphone in terms of inverse probability in order to fasten the algorithm. The results show that the approach efficiently build uniformly distributed phonetically-rich corpus with optimum number of sentences. Keywords—Phonetically-Rich Sentences; Statistical Analysis; Phonemes; Triphone; Hindi Speech Recognition; Grapheme- Phoneme; Hindi Phonology; Phone Like Units(PLU); I. Iඇඍඋඈൽඎർඍංඈඇ When a set of sentences would be called as phonetically- rich set? The answer to the question depends on two statistical properties of phonetic distribution. First one is to know about the characteristic distribution of phonemes which decides the phonetic richnes of a sentence. On the other hand second prop- erty talks about the phonetic resemblance between extracted sentences and language in study. Evidently, this field of study is significantly related to automatic speech recognition (ASR) and speech synthesis(TTS) [1]. The task of extracting phonetically rich and balanced sen- tences involves, analyzing a large corpus and performing the procedure to extract sentences using sentence extraction crite- ria, based on various stochastic methodologies [2]. Addition- ally, a corpus has to be phonetically balanced through made up of sentences having phonetic units as per its distribution in the natural spoken language. Now based on these set of sentences which contains all the phonemes in most of the possible contexts, when recorded, would produce a phonetically rich and balanced speech corpus [3], [4]. Compare to some studies which are based on employing words, syllables and monophones, most of the current research in the development of Automatic Speech Recognition (ASR) and Text to Speech (TTS) systems widely focused on how the contextual phone units e.g. triphones and diphones could be used for improving the robustness of the systems [5], [6]. The field of constructing a phonetically-rich sentence cor- pus is relevant to various applications i.e. ASR and speech synthesis and many more. For instance, to estimate a robust accoustic model, a phonetically rich speech dataset is a basic requirement [7]. Phonologists have found useful to have such a specific corpora in order to develop a sytem to analyze speech production and variability [8]. In speech therapy, to find com- municative disorders of patients, phonetically-rich sentences are often utilized in assessment of patient’s speech production under surveillance in various phonetic/phonological contexts [9]. A novel approach based on the triphone distribution has been formulated and evaluated in this paper. The problem could be elaborated formally as: suppose a corpus C is given which consists of s sentences, find a subset K such that subset K containing sk sentences having uniformly distributed triphones. In first appearance the problem looks simple, but due to inherent high computational time complexity, the task is considered to be approached diligently. The problem is a non-polynomial type as it generates a set of instances based on the combinations of triphones and should be considered as an intractable problem [10]. A two phase algorithm has been proposed to extract set of phonetically-rich triphone sentences from EMILLE corpora in order to build a phonetically-rich corpus for Hindi.

…(Full text truncated)…

This content is AI-processed based on ArXiv data.