Phonemic Dictionary Research Articles

A Systematic Analysis on Wujuh Qira’at in Ablution’s Verses and its Implication to the Islamic Law for Recognition System

Qiraator wujuh qira’at is the methodof recitationof the Quran and the reading varies based on the school of the Imam. The difference within the wujuh qira’atgive effects to the meaning and rules. This article presents the implication of the rules through qira’at mutawatirahby focusing on the ablution verse from the Māidah, verse 6. The research methodlogy adopted is fully based on literature studies with heavy reference on the qiraatbooks, tafsir, hadith, and fiqh through compilation of opinions from AhlusSunnah wal Jamaahscholars. The research findings discovered that varieties in the wujuh qira’atgives more than one meaning for a particular verse. Nonethless, the varieties of the wujuh qira’atdoes not affect the Quranic verse recognition system when referring to the same word. The difference can be detected though the analysis of phonemes that are semantically verified in a phoneme dictionary, hence producing a specific word that carries a specific meaning.

Phonemic status of Bangla nasal vowels: A corpus study

This study investigated the phonemic status of the nasal vowels in Bangla (aka Bengali). It has been claimed for decades that all the seven monophthongal oral vowels in Bangla have phonemically contrastive nasal counterparts; however, an in-depth investigation of the status of nasality for all the vowels is lacking in the current literature. With a phoneme dictionary build from a text corpus of 8 (eight) million word-tokens and about 275 thousand word-types, this study investigated whether all the oral vowels have phonemically contrastive nasal vowels. Findings revealed that only five of the seven monophthongal vowels form phonemically contrastive relationships with their nasal counterparts; nasality in /æ/ and /ɔ/ are not contrastive phonemically.

Urdu Language Translator using Deep Neural Network

Urdu language of Pakistan has more than 100 million speakers in Pakistan, India, Afghanistan and Middle East. With low English literacy rate average Urdu speaking person faces barriers in communicating with foreign people in terms of accessing information, carrying business. This paper proposes an interactive Urdu to English language speech translator using deep Neural Network. ASR module in proposed pipeline is composed of deep neural network and is simpler as compared to traditional ASR which requires complex hand engineering like feature extraction and resources like phoneme dictionary. It was clearly seen that the proposed model shows the high accuracy when the input is recorded audio and it shows poor performance with real time input. While one HTTP request per input transcription produced English translation for Text to Text translation using Python Text Blob library. The final output was achieved with a delay of no more than 30 seconds. Furthermore, we have tested and provided some statistical findings, the result shows that value updating for neural network layer’s bias, standard deviation when Adam optimizer parameters are set as follows: beta1=0.9, beta2=0.9 and learning rate =0.01 meanwhile dropout rate was kept to 5% to offer regularization and observed value for scalar maximum lies between 0 and 0.08. There is a little deviation at 0.05 step, value decreases and afterwards that bias maximum scalar increases with positive values and finally increases exponentially at later stages of training further results are discussed in experiment section respectively. The proposed speech recognition model out performs traditional automatic speech recognition systems in efficiency, simplicity and robustness. Keywords: Deep RNN, Language Translator, N-gram LM, Text Blob Translation, Urdu Language, ASR

64. A Computational Analysis of Verbal Fluency in Schizophrenia

Abstract Background: Verbal fluency tasks are widely used to assess language organization and executive functioning. Meta-analyses of verbal fluency performance in individuals with schizophrenia (SZ) indicate greater impairment on semantic (SF) than phonemic fluency (PF). Because SF depends more on efficient word retrieval within subcategories (clustering) and PF more on shifting between subcategories (switching), fluency impairment has been linked to clustering difficulties in SZ. However, most scoring methods rely on subjective judgments to determine clusters and switches, which reduces reliability. Here, we test a novel computational method to objectively capture clustering, switching, and overall retrieval organization from fluency performance in SZ and demographically matched healthy controls (HC). Methods: PF and SF (“animals”) tasks were administered to SZ and HC. Responses were transcribed from audio recordings into text files for computational analysis using VFClust (Ryan et al, 2013). VFClust utilizes latent semantic analysis (LSA) to quantify the semantic similarity of ordered word pairs as well as a combination of the Cargenie Mellon Phonemic Dictionary (CMUDict) with the Levenshtein distance method to determine phonemic similarity of ordered word pairs. Similarity thresholds were established and applied to these pairwise similarity scores (PSS) to determine cluster or chain inclusion. Clusters included words in which every word pair in that cluster had a PSS above threshold. Chains consisted of words in which PSS of adjacent words were above threshold. Switches were counted when PSS were below threshold, signifying a shift from one category to another. Output metrics for each individual included total words produced, mean size of chains and clusters, number of chains, clusters, and switches, and mean PSS across all ordered word pairs for SF and PF. Results: SZ were impaired compared to HC on SF but not PF. SF deficit was driven by a lower number of chains and clusters produced by SZ compared with HC. Group differences in SF were not driven by any other measures. Negative symptoms were inversely correlated with the total number of words produced and the switch count in SF. Conclusion: Objective, computationally derived fluency metrics captured group differences observed in SF in SZ. Interestingly, the lack of difference in cluster size questions SZ deficits in generating sizable clusters. Instead, limitations in the ability to retrieve a variety of semantically associated words (multiple clusters) seem to underlie performance deficits. Intact switching despite lower number of clusters and chains could indicate increased activation between (conventionally) unrelated words. While altered, semantic relatedness could cascade to thought disorder and abnormal interpersonal interactions, in this study, we only observed a significant negative association with negative symptoms.

Investigation of Effects of Different Synthesis Unit to the Quality of Malay Synthetic Speech

Synthesis unit of a speech synthesizer directly affects the computational load and output speech quality. Generally, phoneme is the best choice to synthesize high quality speech. But it requires the knowledge of language to precisely draw the segmentation of words into phonemes. And it is expensive to compose an accurate phoneme dictionary. In this study, another type of synthesis unit is introduced which is letter. In Malay language, the unit size of letter is smaller than phoneme. And using letter as the synthesis unit could ease a lot of efforts because the context label can be created in fully automatic manner without the knowledge of the language. Four systems have been created and an investigation was done to find out how synthesis unit could affect the quality of synthetic speech. Forty eight listeners were hired to rate the output speech individually and result showed that no obvious difference between the output speech synthesized using different synthesis units. Listening test showed satisfactory result in terms of similarity, naturalness and intelligibility. Synthetic speech with polyphonic label showed increment in intelligibility compared to synthetic speech without polyphonic label. Using letter as the synthesis unit is recommended because it excludes the dependency of linguist and expands the idea of language independent front end text processing.

Method for interactive speech recognition and training

A method for creating word models for a large vocabulary, natural language dictation system. A user with limited typing skills can create documents with little or no advance training of word models. As the user is dictating, the user speaks a word which may or may not already be in the active vocabulary. The system displays a list of the words in the active vocabulary which best match the spoken word. By keyboard or voice command, the user may choose the correct word from the list or may choose to edit a similar word if the correct word is not on the list. Alternately, the user may type or speak the initial letters of the word. Then the recognition algorithm is called again satisfying the initial letters, and the choices displayed again. A word list is then also displayed from a large backup vocabulary. The best words to display from the backup vocabulary are chosen using a statistical language model and optionally word models derived from a phonemic dictionary. When the correct word is chosen by the user, the speech sample is used to create or update an acoustic model for the word, without further intervention by the user. As the system is used, it also constantly updates its statistical language model. The system gets more and more word models and keeps improving its performance the more it is used. The system may be used for connected speech as well as for discrete utterances.

Letter‐to‐phoneme rules: A semi‐automatic discovery procedure

A 20 000‐word phonemic dictionary has been compiled from several sources and hand‐edited so as to contain phonemic representations that are ideal input for the Klattalk synthesis‐by‐rule program or for an automated phonetic alignment system such as is planned for the Spire phonetics work station. The phonemic representation consists of 40 phonemes, primary and secondary stress symbols, and morpheme‐structure symbols to indicate a prefix, suffix, or compound. Computer programs were then written to create letter‐to‐sound rules of the form A→ [b]/CD EF, i.e., the letter “A” goes to the phoneme [b] in the letter environment consisting of left context CD and right context EF. Each instance of the letter “A” in the dictionary creates a rule that may be new, the same as an old rule, or conflicting with an old rule (a letter goes to a different phoneme in the same context). Conflicting rules (which limit the performance of this procedure) are inevitable unless each whole word is stored in the left/right context. Conflicting rules arise because English spelling does not indicate where morpheme boundaries are located, which syllables are stressed, or the fact that the present work has undergone exceptional historical sound change or has been borrowed from another language. If only the most popular of each set of conflicting rules is used, the computer program can form the rules into a tree for rapid execution. The accuracy and tree size of this approach have been investigated as a function of (a) length of the left context, (b), length of the right context, (c) presence or absence of stress symbols in the orthography (which would require stress prediction from text, a problem that we have not studied as yet), and (d) presence or absence of morpheme boundary structure in the orthography. [Research supported in part by an NIH grant.]

Design of a linguistic statistical decoder for the recognition of continuous speech

Most current attempts at automatic speech recognition are formulated in an artificial intelligence framework. In this paper we approach the problem from an information-theoretic point of view. We describe the overall structure of a linguistic statistical decoder (LSD) for the recognition of continuous speech. The input to the decoder is a string of phonetic symbols estimated by an acoustic processor (AP). For each phonetic string, the decoder finds the most likely input sentence. The decoder consists of four major subparts: 1) a statistical model of the language being recognized; 2) a phonemic dictionary and statistical phonological rules characterizing the speaker; 3) a phonetic matching algorithm that computes the similarity between phonetic strings, using the performance characteristics of the AP; 4) a word level search control. The details of each of the subparts and their interaction during the decoding process are discussed.

An overview of the Lincoln Laboratory speech recognition system

The Lincoln speech recognition system is capable of recognizing spoken sentences made up of words drawn from a limited vocabulary and constrained to conform to a context-free grammar. Speech input is taken from a close-talking, noise-cancelling microphone in a relatively noisy computer room. The speech is digitized and subjected to a detailed acoustic-phonetic analysis [C. J. Weinstein et al., Proc. IEEE Symposium on Speech Recognition (April 1974), pp. 89–100] which produces a string of acoustic phonetic elements (APELs). The APEL string is scanned by a linguistic module which attempts to find and score candidate sentences which satisfy the syntactic and semantic constraints of the grammar and which are composed of words having acceptable matches between phonemic dictionary spellings and APEL representations. Typical sentences of 3- to 4-sec duration require the order of one minute of computer processing. Tests have involved vocabularies of 125 to 500 words, grammar of varying complexity, many speakers (both male and female), and several hundred test sentences. [This work was sponsored by the Advanced Research Projects Agency of the Department of Defense.]

Parsing and word matching in the Lincoln Laboratory, speech recognition system

The Lincoln system is designed to recognize sentences conforming to a context-free grammar. Parsing is guided by a heuristic evaluation function which combines individual word scores into parse path scores. The word scores measure the degree of correspondence between phonemic dictionary spellings and the results of an acoustic-phonetic analysis of the input sentence. Word scoring is based on two computer-generated scoring matrices derived from confusion statistics gathered from 113 sentences. The synchronization problem resulting from missing and spurious segments is simplified by first aligning the vowels (sometimes in more than one combination). The effect of phonological rules is handled by insetting optional phonemes in the dictionary and flagging others as possibly missing. Other rules, which are specific to Lincoln's front-end analysis, predict spurious segments and the effect of neighboring semivowels on vowel classification. [This work was sponsored by the Advanced Research Projects Agency of the Department of Defense.]

Representation and Use of Acoustic-Phonetic Knowledge in the HEARSAY System

The HEARSAY speech understanding system uses acoustic-phonetic knowledge at basically four levels: lexical, syllabic, phonemic, and subsegmental. The lexicon is a list of phonemic spellings of all the words of the given language, possibly with alternates. The system recognizes possible word juncture problems from the phonemic dictionary; for example, the juncture between the words “got to” may be realized with one or two stop releases, and the system is sensitive to both possibilities. At a (roughly) syllabic level, the speech is segmented into the three classes of “voiced,” “silent,” and “fricated” (voiced or unvoiced). (In addition, a “voiced” segment is further subdivided if it contains a significant amplitude minimum between significant maxima.) This segmentation, while crude, is very reliable for many first-lever approximations. At the phonemic level, the system matches an expectation derived from the dictionary spelling to the actual utterance; each candidate word is rated as a function of the goodness of this phonemic-phonetic match. At the subsegmental level, a label is assigned to each 10-msec sample of speech. The inventory of labels is intended to correspond to the steady-state portions of the sustained phone types of English. This labeling is used to characterize each 10 msec for the further higher-level processing.