Phonetic Units Research Articles

An evaluation of sentence selection methods on the different phone-sized units for constructing Indonesian speech corpus

Collecting phonetically balanced text corpus is an important step to develop automatic speech recognition and text-to-speech systems. A corpus should have a small number of sentences but contains all phonetic units, such as monophone, triphone, and pentaphone units. There are exist least-to-most greedy algorithm (LTM + Greedy) and its variant to select the minimum sentence set. The variant is on the sentence scoring method, which affect the number of selected sentences. In this paper, we evaluate the sentence scoring methods by Zhang and Suyanto on LTM + Greedy algorithm. The sentence scoring methods are conducted on triphone and pentaphone units on the collection of sentence set. Triphone and pentaphone units have offered higher quality synthesized speech than monophone unit. The dataset of this paper is Indonesian sentences that collected from holy book translation, news, novel, dialog, monologue, and question sentences. Totally 115,489 sentences are used for the experiments. Based on the experiments, LTM + Greedy by Suyanto produces a smaller number of sentences that contain large number of phone units.

Basic auditory processing deficits in infants at risk for dyslexia during the sensitive period predict future language

It is widely agreed that poor reading and spelling performances in individuals with developmental dyslexia underlie problems with processing phonetic units. According to recent child and adult data, such processing deficits likely are the consequence of compromised low-level auditory processing. A possible causal relationship can be explored by testing infants based on their familial risk of dyslexia. Using magnetoencephalography, the study at hand investigates low-level auditory processing in infants at risk for dyslexia across 6 and 12 months of age—a sensitive period in early language development when the auditory system specializes in native phoneme perception. Results showed smaller and shorter neural responses to simple sounds in at-risk infants at 6 than at 12 months, a pattern that was reversed in age-matched typically developing infants. This interaction was significant both when fitting equivalent current dipoles and when using distributed source modeling and localized to left temporal and left frontal brain regions, indicating its potential impact on early language learning. In addition, atypical auditory responses in at-risk infants consistently predicted later language skills, such as syntactic processing and word production, thus identifying a possible biomarker of dyslexia in the infant brain.

Segmentation of speech on phonetic elements for systems of speech information protection

The article is devoted to the development of speech segmentation algorithm on phonetic elements for the synthesis of speech-like signals in speech information protection systems. The main attention is paid to establishing the boundaries of phonetic units of speech, taking into account the influence of this factor on the quality of the synthesized speech by the compilation method. The features of establishing the boundaries of phonemes for fused speech and the influence of this factor on the quality of synthesized speech on the basis of phonemes are considered. It is proposed to ensure the quality of synthesized speech beginning and ending phonemes at the segmentation set in the transition implementation of a signal through zero and in the synthesis of speech-like signals to use the spline function at the boundaries of segments phonemes.

Phonology based Fuzzy Phoneme Recognition

Speech enhanced applications are fundamentally based on an effective separation and recognition of phonetic units. The variable nature of speech signal makes it challenging due to higher rate of variability in the phonetic units. This nature of variability is addressed with the help of Fuzzy Logic in this work for identification of phonemes. The feature parameters representing the speech signal variability in the work are Zero Crossing Rate (ZCR) and Short Term Energy (STE). The approach used in the presented work demonstrate to be more effective than other methods such as Artificial Neural Networks (ANN). A Fuzzy Logic model is tuned using features with various phonemes. Phonology based fuzzy decision logic is used to infer about phonetic units separation point. Phonetic analysis has been done for different spoken words of English language identify the vowel recognition.

Itakura–Saito Divergence as an Element of the Information Theory of Speech Perception

On the basis of the information theory of speech perception, a substantiation of the symmetric form of the Itakura–Saito divergence as the minimum decision-making statistics of the asymptotically optimal algorithm for recognition of speech signals at the basic, phonetic level of their processing is given. The conclusions of the theoretical study are confirmed by the results of an experiment. It is shown that the use of the synthesized algorithm can significantly improve the accuracy and reliability of the automatic recognition of the most problematic phonetic units.

Graphemic-phonetic diachronic linguistic invariance of the frequency and of the Index of Coincidence as cryptanalytic tools.

Languages have inherent characteristics that make them their own and differentiated entities within their phyla and families. Even messages written in any language and later encrypted by cryptographic systems do not lose all of their characteristics, there remain aspects that help the cryptanalyst to recover them without knowing the decryption keys. For the characterization of the languages we will consider the frequencies of their graphemic and phonetic units and the Index of Coincidence, tools of fundamental utility in the field of Cryptography. Their diachronic invariance or survival over time in one language and their ability to discriminate against other languages will be analized. In order to do so, we will examine a total of 101 languages of which 261 texts have been taken. All of them are very diverse in style and time, taking us through a wide linguistic and temporal spectrum that will cover the period from the 6th century BC to the present day.

А Study on Ethnic Minority Languages in Siberia: Focused on Buryat Language

This article is devoted to the study of the Buryаt language. The alphabet of the Buryat language is built on the graphic basis of the Russian language and it consists of 36 letters. Among them, three additional letters were introduced: ү, өө, һ, which have no correspondence in the Russian alphabet. Vowels in the Buryat language are only 12 and are divided into two groups: short and long. The short vowels include: а, о, э, и, у, ү, and long vowels include: аа, оо, ээ, ии, уу, өө, үү, ы. Short vowels are divided into complete vowels and incomplete vowels depending on their location. In the Buryat language, as in all Mongolian languages, there is a certain law of observance of harmony of vowels in a word. In the Buryat word there can be one or several syllables, which are purely phonetic units, not connected by semantics. Regarding stress in the Mongolian languages, there are various judgments: some researchers of the Mongolian languages, in particular, and Buryat, tend to argue about the presence of word stress in them, while others deny its existence.

Development and analysis of multilingual phone recognition systems using Indian languages

In this paper, the development of Multilingual Phone Recognition System (Multi-PRS) using four Indian languages—Kannada, Telugu, Bengali, and Odia—is described. Multi-PRS is an universal Phone Recognition System (PRS), which performs the phone recognition independent of any language. International phonetic alphabets based transcription is used for grouping the acoustically similar phonetic units from multiple languages. Multilingual phone recognisers for Indian languages are studied using two broad groups namely—Dravidian languages and Indo-Aryan languages. Dravidian and Indo-Aryan languages are grouped separately to develop Bilingual PRSs. We have explored both HMMs and DNNs for developing PRSs under both context-dependent and context-independent setups. The state-of-the-art DNNs have outperformed the HMMs. The performance of Multi-PRSs is analysed and compared with that of the monolingual PRSs. The advantages of Multi-PRSs over monolingual PRSs are discussed. Further, we have developed tandem Multi-PRSs using phone posteriors as tandem features to improve the performance of the baseline Multi-PRSs. It is found that the tandem Multi-PRSs have outperformed the baseline Multi-PRSs in all the cases.

Comparison of Phonemic and Graphemic Word to Sub-Word Unit Mappings for Lithuanian Phone-Level Speech Transcription

Conventional large vocabulary automatic speech recognition (ASR) systems require a mapping from words into sub-word units to generalize over the words that were absent in the training data and to enable the robust estimation of acoustic model parameters. This paper surveys the research done during the last 15 years on the topic of word to sub-word mappings for Lithuanian ASR systems. It also compares various phoneme and grapheme based mappings across a broad range of acoustic modelling techniques including monophone and triphone based Hidden Markov models (HMM), speaker adaptively trained HMMs, subspace gaussian mixture models (SGMM), feed-forward time delay neural network (TDNN), and state-of-the-art low frame rate bidirectional long short term memory (LFR BLSTM) recurrent deep neural network. Experimental comparisons are based on a 50-hour speech corpus. This paper shows that the best phone-based mapping significantly outperforms a grapheme-based mapping. It also shows that the lowest phone error rate of an ASR system is achieved by the phoneme-based lexicon that explicitly models syllable stress and represents diphthongs as single phonetic units.

A forced gaussians based methodology for the differential evaluation of Parkinson's Disease by means of speech processing

Abstract Literature evidences the existence of hypokinetic dysarthria in parkinsonian patients and, consequently, the objective characterization of the dysarthric signs associated to the articulatory aspect of speech can be used to detect Parkinson's Disease (PD) providing clinicians with new tools to support the clinical diagnosis. However, no work has analyzed in detail the importance of the different phonemes in the automatic detection of PD from the speech. This work proposes new approaches for this detection by using new classification schemes that allow to compare independently the different phonetic units of patients and controls employed during several speech tasks. Three different parkinsonian corpora were used allowing cross-validation and cross-corpora trials. The results of cross-validation trials (k-folds) provided accuracies between 81% and 94%, with AUC between 0.87 and 0.97 depending on the corpus, while cross-corpora trials yielded accuracies between 66% and 76% with AUC between 0.76 and 0.87. These results suggest that PD affects to the articulatory sequence as a whole, influencing more clearly phonetic units requiring a higher narrowing of the vocal tract. Additionally, text-dependent utterances are considered as the recommended speech task for the detection of PD in this type of schemes as these allow to compare more precisely the phonetic units of patients and controls. Lastly, this work discusses the existence of a glass ceiling in the accuracy of the systems for the automatic detection of PD using speech, concluding that this is below 95% for most of the cases.

A robust unsupervised pattern discovery and clustering of speech signals

In this paper, a novel approach to unsupervised pattern discovery for speech signals is proposed. The proposed work deviates from the standard speech recognition task, and aims to cluster the speech utterances based on the vocabulary of a broad topic. It attempts to discover the matched sequence of phonetic units by making use of the repeated patterns between the speech signals. Identification of matched sequence of phonetic patterns helps in clustering the speech signals, automatically. The proposed approach uses the posterior features derived from Gaussian mixture model (GMM) to find the repeated structure between the speech signals. Image processing techniques are used to identify these matched acoustic patterns. An angle histogram-based method is used to extract the desired matched keyword/phrase patterns present in a pair of speech utterances. The performance of the proposed method is evaluated on Hindi and Bengali news speech corpora using standard objective measures, and also compared with state-of-the-art techniques. The matched pairs of speech utterances obtained by the proposed method are grouped into broader classes using an appropriate clustering technique. The final clusters represent the broader classes of information such as politics, sports, and weather.

Language Recognition using Neural Phone Embeddings and RNNLMs

New advances in Language Identification (LID) using Recurrent Neural Networks (RNNs) and Neural Embeddings have been proposed recently. While their application has been successfully applied at a word level, results at a phoneme level may not be as good because of the greater variability found in phoneme sequences which reduces LID accuracy. Thus, we propose to use phonetic units called “phonegrams” that implicitly include longer-context information and use them to train neural embeddings and RNN language models (RNNLMs). Neural embeddings are used in a pre-processing data phase to reduce the scattering problem produced by the high number of resulting phone-gram units, and, in a second phase, we have used the RNNLMs to obtain the scores of each language in the identification task following a PPRLM structure. Results in terms of Cavg on the KALAKA-3 database show that the use of phone-grams provides up to 14.4% relative improvement over a baseline using only phonemes as features. In addition, our proposed strategy of reducing the number of phone-gram units using neural embeddings contributes to obtain up to 23.0% relative improvement. Finally, fusing the best system with MFCC-based acoustic i-vectors and a traditional PPRLM architecture provides up to 39.3% improvement.

Weighted fast sequential DTW for multilingual audio Query-by-Example retrieval

This paper examines multilingual audio Query-by-Example (QbE) retrieval, utilizing the posteriorgram-based Phonetic Unit Modelling (PUM) approach and the Weighted Fast Sequential Dynamic Time Warping (WFSDTW) algorithm. The PUM approach employs phone recognizers trained on language-specific external resources in a supervised way. Thus, the information about the phonetic distribution is embedded in the process of acoustic modelling. The resulting acoustic models were also used for language-independent QbE retrieval. The improved WFSDTW algorithm was implemented in order to perform retrievals for each query (keyword) within the particular utterance file. The major interest is placed on a retrieval performance measurement of the proposed WFSDTW solution employing posteriorgram-based keyword matching with Gaussian mixture modelling (GMM). Score normalization and fusion of four different language-dependent sub-systems was carried out using a simple max-score merging strategy. The results show a certain predominance of the proposed WFSDTW solution among two other evaluated techniques, namely basic DTW and segmental DTW algorithms. Also, the combination of multiple PUM techniques together with the WFSDTW has been proved as an effective solution for the QbE task.

A comparison of three spectral features for phone recognition in sub-optimal environments

This paper presents a comparison of three spectral features for automatic phone recognition in sub-optimal environments. An exclusive study is carried out with a phone recognition system called phonetic engine (PE) developed in the Manipuri language. The Manipuri language is a scheduled Indian language being used as the official language in the State of Manipur. However, there is no standard database of the language so far. Therefore, a PE has been built for this language. Here phonetic transcriptions are done and then modeling of each phonetic unit is carried out using hidden Markov model (HMM). Speech feature extraction is a very important stage in the development of such a PE. An analysis of phone recognition accuracies of the PE due the three dominant spectral features: MFCC, PLP and LPCC have been studied here. It is found that PLP and MFCC outperform LPCC features under all circumstances.

The influence of the assimilation operator, speech rate and linguistic boundary on the production of /z/ in Croatian

It is widely accepted that invariant and discrete phonological units at the linguistic level are transformed into variable and continuous movements of speech organs, which in turn results in equally continuous acoustical results. The variability of phonemic units depends on neighbouring phonetic units, but also on the various linguistic, communicational and pragmatic contexts of a particular speech act. The influence of phonetic units upon each other results in adaptations, coarticulations and assimilations. By means of assimilation at least one distinctive feature of a phoneme is changed, so the observed phoneme becomes similar to its neighbouring sound – the assimilation operator. This paper is aimed at analysing the influence of speech rate on assimilation processes in the voiced fricative /z/, when it is preceded by sounds /s, z, ʃ, ʒ / in four different types of articulatory joint: sentence, clausal, lexemic and proclitical. The articulatory joint refers to the production of two phonemes separated by different types of linguistic boundaries. Twenty female native speakers of Croatian with no history of speech or hearing impairments read a text at both natural and fast speech rates. The acoustical recording was performed in a sound-treated room. The Praat software was used to analyse six variables in all occurrences of the sound /z/: duration, spectrum centre of gravity, standard deviation of the centre of gravity, spectral skewness, spectral kurtosis, and harmonic to noise ratio. The results showed that various linguistic boundaries, speech rates and sounds as assimilation operators influence the degree of assimilation of the phoneme /z/, as measured by the acoustic variables.

RHETORIC (PROSODY) IN THE LYRICS OF BIDDLE DEHLAVĪ

In the Persian poem, the elements constituting the music are: metrical prosody, the rhymes, and homophony of words in the poem that is the result of the artistic repetition of the various phonetic units of the language. Harmony between these elements and poem content can be considered as a footnote of this phonetic phenomenon and Complementary of the musical quality of the lyrics. In the form of “sonnet”, with regard to the structural characteristics, the mentioned musical elements can be divided into two groups: one is related to the sonnet structure (which is constantly in a sonnet) and other is related to the couplet (which in the different couplet of a sonnet may be changed). The Biddle Dehlavī lyrics is described by “roam” adjective, but his sonnet meter is to a large extent “ordinary”, because high frequency meters of the sonnet, which four fifth of sonnets were versified in that format, are according to the lyric tradition of that time period. However, in low frequency meters of his sonnets, there are uncommon meters that distinguishes it from other poets. Most of the readers of Biddle poems and experts in the Biddle poems believe that his sonnets are the most valuable of his poems. In the other hand, because of its history and lyric nature, the sonnet form has a stronger link to the musics and the meter role is more highlighted in his sonnets.

An analysis of the influence of deep neural network (DNN) topology in bottleneck feature based language recognition.

Language recognition systems based on bottleneck features have recently become the state-of-the-art in this research field, showing its success in the last Language Recognition Evaluation (LRE 2015) organized by NIST (U.S. National Institute of Standards and Technology). This type of system is based on a deep neural network (DNN) trained to discriminate between phonetic units, i.e. trained for the task of automatic speech recognition (ASR). This DNN aims to compress information in one of its layers, known as bottleneck (BN) layer, which is used to obtain a new frame representation of the audio signal. This representation has been proven to be useful for the task of language identification (LID). Thus, bottleneck features are used as input to the language recognition system, instead of a classical parameterization of the signal based on cepstral feature vectors such as MFCCs (Mel Frequency Cepstral Coefficients). Despite the success of this approach in language recognition, there is a lack of studies analyzing in a systematic way how the topology of the DNN influences the performance of bottleneck feature-based language recognition systems. In this work, we try to fill-in this gap, analyzing language recognition results with different topologies for the DNN used to extract the bottleneck features, comparing them and against a reference system based on a more classical cepstral representation of the input signal with a total variability model. This way, we obtain useful knowledge about how the DNN configuration influences bottleneck feature-based language recognition systems performance.

An Analysis of Automatic Phone Recognition and Identification of a Few Languages from North Eastern India

Objective: Phones are basic sound units available in the spoken data. Languages differ among themselves due to use of different phone sets. This paper analyzes some aspects of automatic phone recognition and subsequent identification of a few languages from North Eastern region of India using Phonetic Engine (PE). Methods and Statistical Analysis: PE is a system which converts a speech sample into some symbolic form so that these symbols are capable of capturing all the information carried by the speech sample. In the development of PEs, the International Phonetic Alphabet (IPA) symbols are used in the data transcription process. In modelling the phonetic units Hidden Markov Models (HMM) have been used in the training phase. These trained HMMs are then used in phone recognitions leading to the identification of language(s) of unknown test utterances. Findings: PEs are built for three Indian Languages and one dialect, namely Manipuri, Assamese and Bengali and the Kakching dialect of Manipuri. These languages are widely spoken across the North Eastern region of India. The overall phone recognition accuracies reported by the PEs for the above selected languages are 62:11% for standard Manipuri language, 59:0% for Kakching dialect of Manipuri, 43:28% for standard Assamese and 48:58% for Bengali language. Application: Automatic LID is possible using a set of PEs in testing unknown utterances due to the language bias of these systems. Various level of identification rates reported in some LID tasks carried out with PEs are discussed here to make an analysis of the issues belonging to it.

Эффективность использования метода фоносемантического анализа при переводе японской звукоизобразительной лексики на русский язык (на материале японских поэтических текстов хайку)

The article is devoted to consideration of efficiency of the method of phonosematic analysis in the translation of sound descriptive vocabulary from Japanese into Russian in the most capacious genre of Japanese poetry - haiku. The relevance of the topic is dictated by the lack of clearly defined principles of translation of sound descriptive lexis in general and by the translators ignoring the ability of phonetic units and the sound organization of the text to act as a carrier of emotional content in particular. The article shows that the traditional principles of translating non-equivalent vocabulary, which includes sound descriptive words, are not effective in the case of the Japanese language, mostly because of the peculiarities of Oriental lyric poetry, rooted in the philosophy of Buddhism, Zen (chin. Chan ) and the national world view, which, in turn, is based on imagery and innuendo. On the material of poetic works by the famous Japanese poet Issa Kobayashi (1763-1828), full of onomatopoetical vocabulary, the features of the semantic structure of sound descriptive words are identified, phonetic means are determined, not only relevant for the adequate representation of this structure, but also for ensuring the unity of form and meaning of sound descriptive words.

Heterophonic speech recognition using composite phones.

Heterophones pose challenges during training of automatic speech recognition (ASR) systems because they involve ambiguity in the pronunciation of an orthographic representation of a word. Heterophones are words that have the same spelling but different pronunciations. This paper addresses the problem of heterophonic languages by developing the concept of a Composite Phoneme (CP) as a basic pronunciation unit for speech recognition. A CP is a set of alternative sequences of phonemes. CP’s are developed specifically in the context of Arabic by defining phonetic units that are consonant centric and absorb phonemically contrastive short vowels and gemination, not represented in the Arabic Modern Orthography (MO). CPs alleviate the need to diacritize MO into Classical Orthography (CO), to represent short vowels and stress, before generating pronunciation in terms of Simple Phonemes (SP). We develop algorithms to generate CP pronunciation from MO, and SP pronunciation from CO to map a word into a single pronunciation. We investigate the performance of CP, SP, UG (Undiacritized Grapheme), and DG (Diacritized Grapheme) ASRs. The experimental results suggest that UG and DG are inferior to SP and CP. For the A-SpeechDB corpus with MO vocabulary of 8000, the WER for bigram and context dependent phone are: 11.78, 12.64, and 13.59 % for CP, SP_M (SP from manual diacritized CO), and SP_A (SP from automated diacritized MO) respectively. For vocabulary of 24,000 MO words, the corresponding WER’s are 13.69, 15.08, and 16.86 %. For uniform statistical model, SP has a lower WER than CP. For context independent phone (CI), CP has lower WER than SP.