Abstract

In this report, we present our system for the 4th Computational Linguistics Scientific Document Summarization Shared Task (CLSciSumm 2018), which poses the challenge of identifying the spans of text in a reference paper (RP) that most accurately reflect a citation (i.e. citance) from another paper to the RP and identifying the discourse facet of the corresponding text. We modeled Task1A as Classification Prob-lem which classifies a pair of sentences into classes paraphrase and nonparaphrase. We used Convolutional Neural Network along-with transfer learning for this. Task1B is solved as multi label classification problem.

Abstract

Fricatives are produced by creating a turbulence in the air– flow by passing it through a stricture in the vocal tract cavity. Fricatives are characterized by their noise–like behavior, which makes it difficult to analyze. Difference in the place of articulation leads to different classes of fricatives. Identification of fricative segment boundaries in speech helps in improving the performance of several applications. The present study attempts towards the identification and classification of fricative segments in continuous speech, based on the statistical behavior of instantaneous spectral characteristics. The proposed method uses parameters such as the dominant resonance frequencies, the center of gravity along with the statistical moments of the spectrum obtained using the zero time windowing (ZTW) method. The ZTW spectra exhibits a high temporal resolution and therefore gives accurate segment boundaries in speech. The proposed algorithm is tested on the TIMIT dataset for English language. A high identification rate of 97.5% is achieved for segment boundaries of the sibilant fricative class. Voiced nonsibilants show a lower identification rate than their voiceless counterparts due to their vowel–like spectral characteristics. A high classification rate of 93.2% is achieved be

Abstract

Impulse-like characteristics of excitation occur at the glottal closure instant (GCI) due to sharp closure of the vibrating vocal folds in each glottal cycle. The GCIs are detected from the excitation component of the speech signal, and the excitation component is derived using inverse filtering or its variants. In this paper we propose a method for GCI detection based on single frequency filtering (SFF) of the speech signal. The SFF output has high signal-to-noise ratio (SNR) property in speech regions. The variance (across frequency) contour computed from the SFF output show rapid changes around the GCIs, and these rapid changes can be observed even when the speech signal is degraded. Thus the GCI locations can be extracted even from degraded speech using the SFF analysis. The robustness of the method is demonstrated for several cases of degradation of speech signal.

Abstract

This paper describes the BUT ‘Jilebi’ team’s speech recognition systems created for the 2018 low resource speech recognition challenge for Indian languages. We investigate modifications of multilingual time-delay neural network (TDNN) architectures with transfer learning and compare them to bi-directional residual memory networks (BRMN) and bi-directional LSTM. Our best submission based on system combination achieved word error rates of 13.92% (Tamil), 14.71% (Telugu) and 14.06% (Gujarati). We present the details of submitted systems and also the post-evaluation analysis done for lexicon discovery using unsupervised word segmentation.

Abstract

In this study, classification of the phonation modes in singing voice is carried out. Phonation modes in singing voice can be described using four categories: breathy, neutral, flow and pressed phonations. Previous studies on the classification of phonation modes use voice quality features derived from inverse filtering which lack in accuracy. This is due to difficulty in deriving the excitation source features using inverse filtering from singing voice. We propose to use the excitation source features that are derived directly from the signal. It is known that, the characteristics of the excitation source vary in different phonation types due to the vibration of the vocal folds together with the respiratory effort (lungs effort). In the present study, we are exploring excitation source features derived from the modified zero frequency filtering (ZFF) method. Apart from excitation source features, we also explore cepstral coefficients derived from single frequency filtering (SFF) method for the analysis and classification of phonation types in singing voice. Index Terms: Singing voice, Excitation source, Ph

Abstract

In this paper, we consider breathy to tense voices, which are often considered to be opposite ends of a voice quality continuum. Along with these, other aspects of a speaker's voice play an important role to convey the information to the listener such as mood, attitude and emotional state. The glottal pulse characteristics in different phonation types vary due to the tension of laryngeal muscles together with the respiratory effort. In the present study, we are deriving the features that can capture effects of excitation on the vocal tract system through a signal processing method, called as zero-time windowing (ZTW) method. The ZTW method gives the instantaneous spectrum which captures the changes in the speech production mechanism, providing higher spectral resolution. The cepstral coefficients derived from ZTW method are used for the classification of phonation types. Along with zero-time windowing cepstral coefficients (ZTWCCs), we use the excitation source features derived from zero frequency filtering (ZFF) method. The excitation features used are: strength of excitation, energy of excitation, loudness measure and ZFF signal energy. Classification experiments using ZTWCC and excitation features reveal a significant improvement in the detection of phonation type compared to the existing voice quality features and MFCC features

Abstract

We show an effective way of adding context information to shallow neural language models. We propose to use Subspace Multinomial Model (SMM) for context modeling and we add the extracted i-vectors in a computationally efficient way. By adding this information, we shrink the gap between shallow feed-forward network and an LSTM from 65 to 31 points of perplexity on the Wikitext-2 corpus (in the case of neural 5-gram model). Furthermore, we show that SMM i-vectors are suitable for domain adaptation and a very small amount of adaptation data (e.g. endmost 5 % of a Wikipedia article) brings a substantial improvement. Our proposed changes are compatible with most optimization techniques used for shallow feedforward LMs.

Abstract

This paper focuses on the problem of estimating fundamental frequency from singing voice. Estimation of fundamental frequency is a well studied topic in the speech research community. From the recent studies on fundamental frequency estimation from singing voice with state-of-art methods proposed for speech, there exists a significant gap in accuracy for singing voice. This is mainly because of the wider and rapid variations in pitch in singing voice compared to that in speech. To overcome this, in this paper we propose a method to derive the fundamental frequency from singing voice by exploiting the harmonics of impulse-like excitation in sequence of glottal cycles. The proposed method is compared with the eight state-of-art methods such as YIN, SWIPE, YAAPT, RAPT, SRH, SFF CEP, PEFAC and SHRP on the LYRICS singing database. From the experimental results, it is observed that the accuracy of fundamental frequency by the proposed method is better than many state-of-art methods in various singing categories and laryngeal mechanisms

Abstract

Graph grammars, being a natural extension of string grammars as well as tree grammars, are highly expressive as well as powerful enough to capture the generative structure of a wide variety of scenarios - both simple and complex. But, unfortunately, we do not find many instances which leverage this power graph grammars provide us. Most of the graph grammars available so far are either toy grammars or limited to addressing highly specialized problems. In this paper, we present Vaiśeṣika Graph Grammar (VGG) system, which is as wide as any graph grammar can get i.e., a grammar for reality as theorized in Vaiśeṣika - the foundational ontology found in Indian analytic tradition. This paper extends the work presented in [1]. In [1], the generative as well as the interpretation rules of this system were presented. Here, apart from briefly discussing some of these rules, we also present a parser for such a system which makes it extremely powerful in terms of its ability to classify an input graph as a valid Vaiśeṣika graph or an invalid one which, in itself, is an immensely significant breakthrough for any ontological application. Apart from that, we also present a verification mechanism to verify the interpretations of graphs generated by the generative rules. This is done, at least statistically, if not formally, to give a statistical proof of the soundness of the system (that every generated graph has at least one Vaiśeṣika

Abstract

With neural networks rapidly becoming deeper, there emerges a need for compact models. One popular approach for this is to train small student networks to mimic larger and deeper teacher models, rather than directly learn from the training data. We propose a novel technique to train studentteacher networks without directly providing label information to the student. However, our main contribution is to learn how to learn from the teacher by a unique strategy - having the student compete with a discriminator.