The current popular knowledge distillation (KD) methods effectively compress the transformer-based end-to-end speech recognition model. However, existing methods fail to utilize complete information of the teacher model, and they distill only a limited number of blocks of the teacher model. In this study, we first integrate a task-driven loss function into the decoder’s intermediate blocks to generate task-related feature representations. Then, we propose an attention-based multi-level feature distillation to automatically learn the feature representation summarized by all blocks of the teacher model. Under the 1.1M parameters model, the experimental results on the Wall Street Journal dataset reveal that our approach achieves a 12.1% WER reduction compared with the baseline system.

To guide practitioners and researchers to design and research Machine Learning (ML) system development processes, we conduct a preliminary literature review on ML system development practices. We identified seven papers and two other papers determined in an ad-hoc review. Our findings include emphasized phases in ML system developments, frequently described ML-specific practices, and tailored traditional practices.

Photoglottography (PGG) is an effective method to monitor vocal fold vibrations via measuring light transmission across the glottis. The difficulty in operation however limits its wide use in speech studies. This paper is to realize a portable PGG (P-PGG) module with an audio interface to record glottal and speech waveforms simultaneously with ease. Near Infrared (NIR) LEDs are driven as a light source and an extremely high-gain photodetector circuit is employed. The output PGG signal is subsequently band-pass filtered and amplified for recording. The whole system is minimized, battery powered with well-controlled heat radiation. In experiments, P-PGG, EGG and microphone are worn together by speakers. The results verify that the NIR lighting P-PGG is successful at recording complete information of glottal cycles in comparison to EGG. Thus, the P-PGG is an effective method for investigating phonation types and consonant-vowel interactions in speech.

Business-to-Business (B-to-B) software development companies develop services to satisfy their customers’ requirements. Developers should prioritize customer satisfaction because customers greatly influence agile software development. However, satisfying current customer’s requirements may not fulfill actual users or future customers’ requirements because customers’ requirements are not always derived from actual users. To reconcile these differences, developers should identify conflicts in their strategic plan. This plan should consider current commitments to end users and their intentions as well as employ a data-driven approach to adapt to rapid market changes. A persona models an end user representation in human-centered design. Although previous works have applied personas to software development and proposed data-driven software engineering frameworks with gap analysis between the effectiveness of commitments and expectations, the significance of developers’ commitment and quantitative decision-making are not considered. Developers often do not achieve their business goal due to conflicts. Hence, the target of commitments should be validated. To address these issues, we propose Data-Driven Persona Retrospective (DDR) to help developers plan future releases. DDR, which includes the Persona Significance Index (PerSI) to reflect developers’ commitments to end users’ personas, helps developers identify a gap between developers’ commitments to personas and expectations. In addition, DDR identifies release situations with conflicts based on PerSI. Specifically, we define four release cases, which include different situations and issues, and provide a method to determine the release case based on PerSI. Then we validate the release cases and their determinations through a case study involving a Japanese cloud application and discuss the effectiveness of DDR.

Vocal tract resonances give rise to core spectral information of speech signals. Linear prediction and cepstral methods are widely used for this purpose. However, both approaches are prone to fail as the fundamental frequency (F0) rises. In this study, a new cepstral method is developed combined with a refined rahmonic subtraction technique (RS-CEPS) to extract spectral envelopes excited by glottal noise sources. A vowel synthesis system based on 3D-printed solid vocal tract models is used to obtain reference transfer functions for accuracy verification. A series of stable vowels /a/ was synthesized for a wide F0 range. By analyzing the synthetic vowels, the results showed that the RS-CEPS yields accurate estimates of resonance-peak and anti-resonance frequencies in comparison to those from the conventional methods. The RS-CEPS is simple and stable, offering a potential for expanding speech analysis applications.

S.1391-1395

Stop consonants in Mandarin Chinese are all voiceless at word-initial positions only showing aspirated and unaspirated distinctions. Between the two phonation types, voice onset time (VOT) shows a clear contrast in duration, whereas voice onset fundamental frequency (onset F0) does not, as seen in previous studies. This study reports our first use of improved instrumentation techniques to record speech sound, oral airflow and glottal activity to examine consonant aspiration and onset F0. Glottal abduction, adduction and vibration cycles are monitored by a new external photo-glottographic system (ePGG) refined for better signal quality to detect accurate voice onset. Experimental data on Mandarin stops obtained from two male subjects suggests that consonant aspiration results in a large variation of VOT and oral airflow at voice onset. The onset F0 shows individual variation of falling and rising contours, and it is higher in aspirated stops than in unaspirated ones.

Application Programming Interfaces (APIs) are common in software development to reuse other products. Although the documentation allows API consumers to learn about API usages, it can be unreliable. Here, we investigate the characteristics of inappropriate usage examples in web API documentation by extracting and comparing OpenAPI Specifications from usage example-response pairs. About 65.5% of the endpoints have some form of inappropriate usage examples. Furthermore, mismatches are classified into four categories: undocumented keys pattern, dynamic keys pattern, unreturned keys pattern, and type mismatched pattern. Our results suggest that the number of keys in the response is correlated with the number of mismatches. These findings should assist both API providers and consumers who deal with unreliable documentation in web APIs.

The importance of APIs in software development, especially web APIs, has increased Developers read documentation, which is available on the internet, and use the corresponding APIs in their products. However, documentation occasionally contains mistakes. Such mistakes can confuse developers or lead to defects that lower the quality of the product. In this paper, we investigate the reliability of web APIs by extracting and comparing OpenAPI specifications from both the documentations and the results of the API calls. Almost half of the documentations are somehow unreliable. Mismatches between documentation and the response can be categorized into four types: 1) Undocumented Keys, 2) Dynamic Keys, 3) Unreturned Keys, and 4) Type Mismatched. This study will help developers design more reliable products.

Today’s development environment has changed drastically; the development periods are shorter than ever and the number of team members has increased. Consequently, controlling the activities and predicting when a development will end are difficult tasks. To adapt to changes, we propose a generalized software reliability model (GSRM) based on a stochastic process to simulate developments, which include uncertainties and dynamics such as unpredictable changes in the requirements and the number of team members. We assess two actual datasets using our formulated equations, which are related to three types of development uncertainties by employing simple approximations in GSRM. The results show that developments can be evaluated quantitatively. Additionally, a comparison of GSRM with existing software reliability models confirms that the approximation by GSRM is more precise than those by existing models.

Feedback control mechanisms for speaking have been examined using the transformed auditory feedback (TAF) technique. Previous studies have shown that speakers demonstrate fundamental frequency (F0) changes when they monitor their voice with artificial alterations of F0. However, those studies underestimate the role of vibrotactile information involved in feedback F0 control. This pilot study aims at exploring whether and how vibrotactile information from the larynx influences vowel F0. Participants in our experiment were asked to sustain vowel with their F0 adjusted to composite sinewave stimuli, which were given via auditory and vibrotactile channels using a headset on the ears or a bone-conduction transducer on the larynx. Results revealed the greater compensatory responses to combined vibrotactile-auditory stimuli than to the responses to auditory-only stimuli. The effect of vibrotactile stimuli on feedback F0 adjustment was also observed with the shorter latency of the responses.

Development environments have changed drastically, development periods are shorter than ever and the number of team members has increased. These changes have led to difficulties in controlling the development activities and predicting when the development will end. Especially, quality managers try to control software reliability and project managers try to estimate the end of development for planning developing term and distribute the manpower to other developments. In order to assess recent software developments, we propose a generalized software reliability model (GSRM) based on a stochastic process, and simulate developments that include uncertainties and dynamics. We also compare our simulation results to those of other software reliability models. Using the values of uncertainties and dynamics obtained from GSRM, we can evaluate the developments in a quantitative manner. Additionally, we use equations to define the uncertainty regarding the time required to complete a development, and predict whether or not a development will be completed on time. We compare GSRM with an existing model using two old actual datasets and one new actual dataset which we collected, and show that the approximation curve generated by GSRM is about 12% more precise than that generated by the existing model. Furthermore, GSRM can narrow down the predicted time range in which a development will end to less than 40% of that obtained by the existing model.

It is important for project stakeholders to identify the states of projects and quality of products. Although metrics are useful for identifying them, it is difficult for project stake-holders to select appropriate metrics and determine the purpose of measuring metrics. We propose an approach that defines the measured metrics by GQM method to identify tendency in projects and products based on Trend Pattern. Additionally, we implement a tool as a Jenkins Plug in to visualize an evaluation results based on GQM method. We perform an industrial case study, which objects are two software development projects. In our industrial case study, we can identify the problem that product contains. As our future work, we will adopt our approach and GQM Plug in to software development project continuously to assess their effectiveness in the long term.

It is important for project stakeholders to identify the states of projects and quality of products. Although metrics are useful for identifying them, it is difficult for project stakeholders to select appropriate metrics and determine the purpose of measuring metrics. We propose an approach that defines the measured metrics by GQM method, and supports identifying tendency in projects and products based on Trend Pattern. Additionally, we implement a tool as a Jenkins Plug in which to visualizes an evaluation results based on GQM method. We perform an experiment with OSS and industrial case study with two software development projects. In our experiment, we can identify the problem and project tendency. In our industrial case study, we can also identify the problem that project contains. As our future work, we will adopt our approach and GQM Plug in to software development project continuously to assess their effectiveness in the long term.

Development environments have changed drastically, development periods are shorter than ever and the number of team members has increased. Especially in open source software (OSS), a large number of developers contribute to OSS. OSS has difficulties in predicting or deciding when it will be released. In order to assess recent software developments, we proposed a generalized software reliability model (GSRM) based on a stochastic process, and compared GSRM with other models. In this paper, we focus on the release dates of OSS and the growth of faults (issues).

Monitoring of water consumption in irrigation systems has become increasingly important for water managers in the actual trend of integrated water management. Low spatial resolution (LSR) satellite remote sensing has already proven the capacity of monitoring evapotranspiration (ETa) over large areas at high temporal frequencies, by which monitoring for large irrigation systems can be satisfied. However, smaller pixel size is still required for more local management while keeping the return period within a few days. High spatial resolution (HSR) satellite imagery is indeed available for calculation of ETa and has already been used in many studies. However, its practical return period is a major drawback to its implementation for monitoring irrigation systems. This paper is perusing into the use of genetic algorithms to assimilate parameters of an agrohydrological model called soil-water-air-plant for each of the pixels of HSR images contained into one single pixel of an LSR multitemporal image. The methodology developed and experimented here is trying to take advantage of the spatial content of HSR images and the temporal content of LSR images by fusing them by the process of data assimilation

Vocal tract data from 3D cine-MRI are used together with synchronised acoustics to evaluate a linear regression model for inversion. The first two principal components of vocalic area functions are predicted with correlations 0.99 and 0.97 respectively, from 24 FFT-cepstra measured in the frequency band 0-4 kHz. This best regression model together with the two component representation yields mean absolute errors of 0.37 cm 2 in section area and 0.15 cm in vocal tract length.

This work proposes an experimental method for direct measurement of transfer functions of acoustic tubes. The method obtains a pressure-to-velocity transfer function from measurement of input volume velocity and output pressures of a target tube. Steady sinusoidal waves from 100 Hz to 5 kHz with a 10-Hz increment were used as a source signal. Experimental results compared with transmission line simulations indicate the following: (1) transfer functions obtained from the measurements agree well with those from transmission line simulations; (2) differences between the resonant frequencies obtained from the measurements and simulations with a uniform tube are less than 2.6 %. These results show conclusive evidence that the proposed method permits accurate measurements of transfer functions of acoustic tubes.

Physiological models of the speech organs must consider cocontraction of the muscles, a common phenomenon taking place during articulation. This study investigated cocontraction of the tongue muscles using the physiological articulatory model that replicates midsagittal regions of the speech organs to simulate articulatory movements during speech [1,2]. The relation between the muscle force and tongue movement obtained by the model simulation indicated that each muscle drives the tongue towards an equilibrium position (EP) corresponding to the magnitude of the activation forces. Contributions of the muscles to the tongue movement were evaluated by the distance between the equilibrium positions. Based on the EPs and the muscle contributions, an invariant mapping (the EP map) was established to function the connection of a spatial location to a muscle force. Cocontractions between agonist and antagonist muscles were simulated using the EP maps. The simulations demonstrated that coarticulation with multiple targets could be compatibly realized using the co-contraction mechanism. The implementation of the co-contraction mechanism enables relatively independent control over the tongue tip and body.

This report describes the development of a physiologically-based articulatory model, which consists of the tongue, mandible, hyoid bone and vocal tract wall. These organs are represented in a quasi-3D shape to replicate a midsagittal layer with a thickness of 2 cm for tongue tissue and 3 cm for tract wall. The geometry of these organs and muscles are extracted from volumetric MR images of a male speaker. Both the soft and rigid structures are represented by mass-points and viscoelastic springs for connective tissue, where the springs for bony organs are set to extremely large stiffness. This design is suitable to compute soft tissue deformations and rigid organ displacements simultaneously using a single algorithm, and thus reduces computational complexities of the simulation. A novel control method is developed to produce dynamic actions of the vocal

Quantitative models of intonation in a variety of languages typically specify a long-range downtrend across the sentence that provides a declining backdrop for the steeper rises and falls of more local pitch events such as accents and word tones.Several studies of this "declination" in English and several other languages have isolated a component of somewhat steeper decline that covers only the last few centiseconds of "lab speech" utterances.Other studies suggest that this "final lowering" may be particular to utterances with a "declarative intonation" pattern, and that it is associated particularly with the ends of discourse units.Thus final lowering seems to be associated pragmatically with a sense of fading off or finality.To see whether final lowering can be attributed in part to a fading off of subglottal pressure, we examined the two measures together in two databases of utterances that varied in intonation contour.To minimize confounds from more local pitch specifications, we looked at the relationship between final lowering and subglottal pressure only in the intonational "tail" -i.e., the portion of the contour after the last pitch accent.The slope of the decline of the subglottal pressure varied as a function of the phonological specification of the tones in the tail.Utterances with declarative intonation or with any other contour sharing the phonological specification of a low tone at the end of the tail consistently showed a decline in subglottal pressure, whereas utterances with "yes-no question intonation" or any other contour sharing the phonological specification of a final high tone showed lesser declines or even increases.

The vowel space reflects the right-angled shape of the vocal tract, and many consonants exploit the palatal wall.These two facts suggest the importance of the geometry of peripheral structure in speech production.In this study, the relationship between geometry and articulatory variation was examined using a database of English and Japanese speakers.The geometry of each speaker's vocal tract was defined by a quadrilateral bounded by the palatal plane and other rigid structures.This quadrilateral, whose area we refer to as the articulatory (or A) space, provides indices of pharyngeal distance, lower facial height, mandibular position and inclination, and head rotation.The A-spaces of different speakers vary in size and form: the speakers with longer pharyngeal distance tend to have shorter lower facial height.There is also significant variation among speakers in the degree of inclination of the mandibular symphysis.Qualitative comparisons suggested that speakers' vowel articulations adapt to the form of their respective A-space, while consonant articulations seem to be independent of the A-space.

A new method for the high-speed digital recording of the vocal fold vibration is presented. The method employs a solid endoscope, a solid-state image sensor and a digital image memory. About 100 frames of continuous image data with 50 × 50 picture elements can be stored in the image memory at one time. Image recording at a rate of 2000 frames per second was achieved using a small light source of a 250 W halogen lamp. The method enables to record the speech signal without special consideration on the camera noises, and also to observe the recorded image on the spot through CRT display. An example of the image data with simultaneously recorded speech and EEG signals are presented comparing the male and female voices, and the chest voice, falsetto and breathy voice.