Corpus Of Spontaneous Japanese Research Articles

Recognition of target domain Japanese speech using language model replacement

End-to-end (E2E) automatic speech recognition (ASR) models, which consist of deep learning models, are able to perform ASR tasks using a single neural network. These models should be trained using a large amount of data; however, collecting speech data which matches the targeted speech domain can be difficult, so speech data is often used that is not an exact match to the target domain, resulting in lower performance. In comparison to speech data, in-domain text data is much easier to obtain. Thus, traditional ASR systems use separately trained language models and HMM-based acoustic models. However, it is difficult to separate language information from an E2E ASR model because the model learns both acoustic and language information in an integrated manner, making it very difficult to create E2E ASR models for specialized target domain which are able to achieve sufficient recognition performance at a reasonable cost. In this paper, we propose a method of replacing the language information within pre-trained E2E ASR models in order to achieve adaptation to a target domain. This is achieved by deleting the “implicit” language information contained within the ASR model by subtracting the source-domain language model trained with a transcription of the ASR’s training data in a logarithmic domain. We then integrate a target domain language model through addition in the logarithmic domain. This subtraction and addition to replace of the language model is based on Bayes’ theorem. In our experiment, we first used two datasets of the Corpus of Spontaneous Japanese (CSJ) to evaluate the effectiveness of our method. We then we evaluated our method using the Japanese Newspaper Article Speech (JNAS) and CSJ corpora, which contain audio data from the read speech and spontaneous speech domain, respectively, to test the effectiveness of our proposed method at bridging the gap between these two language domains. Our results show that our proposed language model replacement method achieved better ASR performance than both non-adapted (baseline) ASR models and ASR models adapted using the conventional Shallow Fusion method.

Toward processing of prosody in spontaneous Japanese

This paper considers how prosody in spontaneous Japanese is processed. We have conducted Rapid Prosody Transcription (RPT) perception experiments on the Corpus of Spontaneous Japanese (CSJ) and investigated how boundaries and prominences are perceived. We recruited three groups of participants from different Japanese dialects and found that (i) F0 is not a strong prominence cue in Japanese, contra Japanese literature on focus prominence (Pierrehumbert & Beckman (P&B) 1988; Kori 1989; Ishihara 2016) and (ii) Japanese allows multi-headed and headless intonation phrases, and P&B’s reset theory, i.e. focal prominence resets boundary phrases, faces empirical difficulties. We also found that (iii) both content words and function morphemes get highlighted in Japanese, and (iv) perception strategies vary cross-dialectally and listeners from different dialects perceive boundaries and prominences differently.

Multiple Sources of Surprisal Affect Illusory Vowel Epenthesis.

Illusory epenthesis is a phenomenon in which listeners report hearing a vowel between a phonotactically illegal consonant cluster, even in the complete absence of vocalic cues. The present study uses Japanese as a test case and investigates the respective roles of three mechanisms that have been claimed to drive the choice of epenthetic vowel—phonetic minimality, phonotactic predictability, and phonological alternations—and propose that they share the same rational goal of searching for the vowel that minimally alters the original speech signal. Additionally, crucial assumptions regarding phonological knowledge held by previous studies are tested in a series of corpus analyses using the Corpus of Spontaneous Japanese. Results show that all three mechanisms can only partially account for epenthesis patterns observed in language users, and the study concludes by discussing possible ways in which the mechanisms might be integrated.

Sociolinguistic Factors Affecting Vowel Devoicing in Spontaneous Japanese: A Preliminary Corpus-based Analysis

Abstract Maekawa and Kikuchi (2005) used the Corpus of Spontaneous Japanese (CSJ) to analyse the frequency of vowel devoicing in different phonological environments. According to their analysis, the devoicing rate is highest when a fricative is followed by a stop and lowest when an affricate is followed by a fricative. Moreover, the results of their study suggest that devoicing also occurs in atypical environments, as in non-close vowels and in contexts where a vowel is followed by a voiced consonant. However, the frequency of devoicing is conditioned not only by phonological factors but, at a certain extent, also by extra-linguistic and sociological factors such as age, gender and speech style. This paper aims to analyse the sociolinguistic variation of vowel devoicing in spontaneous Japanese using the CSJ-Core consisting of about 45 hours of speech, all of which have been (sub-)phonemically segmented. Age, gender and speech style variation has been analysed for different phonological environments. Particular attention will be given to atypical environments which showed a higher rate of variability.

Phonetic variability of nasals and voiced stops in Japanese

We investigated the phonetic variability of nasals and voiced stops in a large-scale Japanese speech corpus. We then analyzed the types of variation across speech styles. In particular, we examined the instances where target segments are deleted or realized as different phonemes. We identified 285 lexical entries displaying variability of target segments in the Corpus of Spontaneous Japanese (Maekawa, 2003). In line with the findings of Arai (1999), we observed a few variants of voiced stops, such as /d/ becoming [n] in /doɴna/ “what” and /g/ being deleted in /daigaku/ “university.” We also found /b/ turning into [m], such as /boku wa/ “I am” becoming [mo kaː]. For nasals, we observed /zeɴiɴ/ “all the people” and /geɴiɴ/ “reason” becoming [zẽːɪɴ] and [gẽːɪɴ], in which the first /ɴ/ was deleted and the preceding /e/ was nasalized and lengthened. Our findings suggest that the extent to which speakers produce phonetic variants of target segments is likely specified lexically more than stylistically because w...

Effects of Surprisal and Entropy on Vowel Duration in Japanese.

Research on English and other languages has shown that syllables and words that contain more information tend to be produced with longer duration. This research is evolving into a general thesis that speakers articulate linguistic units with more information more robustly. While this hypothesis seems plausible from the perspective of communicative efficiency, previous support for it has come mainly from English and some other Indo-European languages. Moreover, most previous studies focus on global effects, such as the interaction of word duration and sentential/semantic predictability. The current study is focused at the level of phonotactics, exploring the effects of local predictability on vowel duration in Japanese, using the Corpus of Spontaneous Japanese. To examine gradient consonant-vowel phonotactics within a consonant-vowel-mora, consonant-conditioned Surprisal and Shannon Entropy were calculated, and their effects on vowel duration were examined, together with other linguistic factors that are known from previous research to affect vowel duration. Results show significant effects of both Surprisal and Entropy, as well as notable interactions with vowel length and vowel quality. The effect of Entropy is stronger on peripheral vowels than on central vowels. Surprisal has a stronger positive effect on short vowels than on long vowels. We interpret the main patterns and the interactions by conceptualizing Surprisal as an index of motor fluency and Entropy as an index of competition in vowel selection.

『日本語話し言葉コーパス』を用いたdephrasing生起要因の分析 : 修飾関係及びモーラ数の効果

『日本語話し言葉コーパス』を用いたdephrasing生起要因の分析 : 修飾関係及びモーラ数の効果

Durational compensation within a CV mora in spontaneous Japanese: Evidence from the Corpus of Spontaneous Japanese.

Previous experimental studies showed that in Japanese, vowels are longer after shorter onset consonants; there is durational compensation within a CV-mora. In order to address whether this compensation occurs in natural speech, this study re-examines this observation using the Corpus of Spontaneous Japanese. The results, which are based on more than 200 000 CV-mora tokens, show that there is a negative correlation between the onset consonant and the following vowel in terms of their duration. The statistical significance of this negative correlation is assessed by a traditional correlation analysis as well as a bootstrap resampling analysis, which both show that it is unlikely that the observed compensation effect occurred by chance. The compensation is not perfect, however, suggesting that it is a stochastic tendency rather than an absolute principle. This paper closes with a discussion of potential factors that may interact with the durational compensation effect.

Maximum-<italic>a-Posteriori</italic>-Based Decoding for End-to-End Acoustic Models

This paper presents a novel decoding framework for acoustic models (AMs) based on end-to-end neural networks (e.g., connectionist temporal classification). The end-to-end training of AMs has recently demonstrated high accuracy and efficiency in automatic speech recognition (ASR). When using the trained AM in decoding, although a language model (LM) is implicitly involved in such an end-to-end AM, it is still essential to integrate an external LM trained with a large text corpus to achieve the best results. While there is no theoretical justification, most of the studies suggest using a naive interpolation of the end-to-end AM score and the external LM score, empirically. In this paper, we propose a more theoretically sound decoding framework derived from a maximization of the posterior probability of a word sequence given an observation. As a consequence of the theory, the subword LM is newly introduced to seamlessly integrate the external LM score with the end-to-end AM score. Our proposed method can be achieved by a small modification of the conventional weighted finite-state transducer-based implementation, without having to heavily increase the graph size. We tested the proposed decoding framework on ASR experiments with the Corpus of the Wall Street Journal and the Corpus of Spontaneous Japanese. The results showed that the proposed framework achieved significant and consistent improvements over the conventional interpolation-based decoding framework.

Revisiting articulatory positions of Japanese vowels as a function of duration on the basis of analysis of large-scale speech corpora

The purpose of this study is to illustrate certain features of the place of articulation of Japanese vowels on the basis of formant frequency. For this purpose, two types of relationship between vowel duration and formant frequency were examined. The first of these relationships was the correlation between speaking rate and formant frequency. Speech materials are obtained from two large-scale speech corpora: Electrotechnical Laboratory Spoken Word database (ETL-WD) and The corpus of spontaneous Japanese (CSJ). The results of analyses showed that as speaking rate slows, articulations get more distinct: open vowels increase in openness, close vowels increase in closeness, front vowels increase in frontness, and back vowels increase in backness. The second relationship was between formant frequencies and vowel length. The result of this analysis showed that place of articulation is more distinct in long vowels than in short vowels. The exception was the back vowel /u/, all of which did not show greater backn...

Age estimation in Japanese speech based on feature selection

This paper addresses age estimation of Japanese speech utterance, using paralinguistic information in speech. Our speech conveys not only linguistic information but also paralinguistic and non-linguistic information. Nonlinguistic information includes potentially valuable social information such as personal parameters, body conditions, gender and ages. This significant information, however, has not been investigated enough. The main purpose of our study was to develop less-invasive methods to extract this information, by proposing a method to classify speakers’ age into five groups (20s, 30s, 40s, 50s, and 60s). 6375-dimentional audio features were extracted from 1579 samples of 913 male speakers extracted from CSJ (Corpus of Spontaneous Japanese) and feature selection was conducted based on Fisher's Ratio. Naive Bayes with top 20 parameters of Fischer’s ratio, yields highest accuracy rate of 51.2%.

Speaker age effects on the voicing contrast of Tokyo Japanese stops

While older Japanese speakers from the Kansai region produce voiced-voiceless stop contrasts with a true voicing distinction, older speakers from the northern Tohoku region produce the same stop contrasts with a short- versus long-lag VOT distinction, similar to the voicing contrast of English stops. However, Japanese speakers of younger generations from both of these dialect regions have been observed to produce the voiced-voiceless stop contrast in much the same way to each other, and they also seem to be using VOT as a less informative cue for stop distinction than speakers of previous generations (Takada, 2011). This study investigates the status of this sound change in the speech from the Tokyo dialect region, generally considered the standard for the modern language. Using speech data available from the Corpus of Spontaneous Japanese, we analyze stops produced by Tokyo speakers to determine differences in voicing contrast conditioned by speaker age and gender. We expect the stop contrast of older sp...

Optimization of the Verbal Inflectional Paradigm by the Cyclic Application of Morphophonological Processes: Evidence from Potential Forms in Japanese

AbstractIn Japanese, a series of morphophonological changes that reorganizes the verbal inflectional paradigms is currently underway. The changes in potential forms involve two innovative processes: ar-Deletion and re-Insertion. This paper analyzes these morphophonological changes in the Japanese potential forms, and models the mechanism that governs the sequential changes, using the Corpus of Spontaneous Japanese. By positing a chronological order between the changes, I argue that the initial change in consonant-final verbs is motivated by semantic disambiguation, and the subsequent change in vowel-final verbs is triggered by analogical leveling for the optimization of the conjugation paradigm. This analogical leveling reorganizes the paradigm in order to counter the increased discrepancy between potential forms of consonant-final verbs and vowel-final verbs caused by the previous change. I propose that positing this two-step process is crucial to understanding the diachronic changes at issue.

An analysis of the singleton-geminate contrast in Japanese fricatives and stops

Previous acoustic analyses of the singleton-geminate contrast in Japanese have focused primarily on read speech. The present study instead analyzed the lengths of singleton and geminate productions of word-medial fricatives and voiceless stops in spontaneous monologues from the Corpus of Spontaneous Japanese (Maekawa, 2003). The results of a linear mixed effects regression model mirrored previous findings in read speech that the geminate effect (the durational difference between geminate and singletons) of stops is significantly larger than that of fricatives. This study also found a large range of variability in the geminate effect size between talkers. The size of the geminate effect between fricatives and voiceless stops was found to be slightly correlated, suggesting that they might be related to other rate-associated production differences between individuals. This suggestion was evaluated by exploring duration differences associated with talker age and gender. While there was no relationship between age and duration, males produced shorter durations than females for both fricatives and stops. However, the size of the geminate effect was not related to the gender of the speaker. The cause of these individual differences may be related to sound perception. Future research will investigate the cause of these individual differences in geminate effect size.

Distributional Pattern of Compound Accent Observed in a Japanese Accent Dictionary and the Corpus of Spontaneous Japanese

Distributional Pattern of Compound Accent Observed in a Japanese Accent Dictionary and the Corpus of Spontaneous Japanese

Prior-shared feature and model space speaker adaptation by consistently employing map estimation

The purpose of this paper is to describe the development of a speaker adaptation method that improves speech recognition performance regardless of the amount of adaptation data. For that purpose, we propose the consistent employment of a maximum a posteriori (MAP)-based Bayesian estimation for both feature space normalization and model space adaptation. Namely, constrained structural maximum a posteriori linear regression (CSMAPLR) is first performed in a feature space to compensate for the speaker characteristics, and then, SMAPLR is performed in a model space to capture the remaining speaker characteristics. A prior distribution stabilizes the parameter estimation especially when the amount of adaptation data is small. In the proposed method, CSMAPLR and SMAPLR are performed based on the same acoustic model. Therefore, the dimension-dependent variations of feature and model spaces can be similar. Dimension-dependent variations of the transformation matrix are explained well by the prior distribution. Therefore, by sharing the same prior distribution between CSMAPLR and SMAPLR, their parameter estimations can be appropriately regularized in both spaces. Experiments on large vocabulary continuous speech recognition using the Corpus of Spontaneous Japanese (CSJ) and the MIT OpenCourseWare corpus (MIT-OCW) confirm the effectiveness of the proposed method compared with other conventional adaptation methods with and without using speaker adaptive training.

A Corpus‐based Investigation of Fillers used by Japanese Native Speakers: -Mainly analysis of Corpus of Spontaneous Japanese-(Part 1)

A Corpus‐based Investigation of Fillers used by Japanese Native Speakers: -Mainly analysis of Corpus of Spontaneous Japanese-(Part 1)

Committee-Based Active Learning for Speech Recognition

We propose a committee-based method of active learning for large vocabulary continuous speech recognition. Multiple recognizers are trained in this approach, and the recognition results obtained from these are used for selecting utterances. Those utterances whose recognition results differ the most among recognizers are selected and transcribed. Progressive alignment and voting entropy are used to measure the degree of disagreement among recognizers on the recognition result. Our method was evaluated by using 191-hour speech data in the Corpus of Spontaneous Japanese. It proved to be significantly better than random selection. It only required 63h of data to achieve a word accuracy of 74%, while standard training (i.e., random selection) required 103h of data. It also proved to be significantly better than conventional uncertainty sampling using word posterior probabilities.

A Quantitative Analysis of Nominative/Genitive Alternation in Japanese

In lieu of an abstract, here is a brief excerpt:Based on his grammaticality judgment test, Harada (1971) claimed that there is an ongoing change, whereby the speakers of Tokyo Japanese increasingly prefer ga to no in relevant environments. In this paper, I present a quantitative analysis using the Corpus of Spontaneous Japanese. The goal of this study is not only to validate the findings in Nambu (2007) and Nambu and Matsuda (2007), but also to give another perspective on the variation.

Topic tracking language model for speech recognition

In a real environment, acoustic and language features often vary depending on the speakers, speaking styles and topic changes. To accommodate these changes, speech recognition approaches that include the incremental tracking of changing environments have attracted attention. This paper proposes a topic tracking language model that can adaptively track changes in topics based on current text information and previously estimated topic models in an on-line manner. The proposed model is applied to language model adaptation in speech recognition. We use the MIT OpenCourseWare corpus and Corpus of Spontaneous Japanese in speech recognition experiments, and show the effectiveness of the proposed method.