Despite all environmental and economic advantages of wind power, noise emission remains an issue for population acceptance. In France, the current noise emission regulation defines noise emergence level thresholds, leading to wind turbine curtailment. Great energy generation losses and thus lost revenues are at stake. This master thesis presents current acoustic campaigns conducted for the development of a wind power project in France and proposes acoustic model improvements to predict curtailment losses before the construction of the wind farm. It first gives insights about the French wind power context and a literature review of available technologies to reduce noise emission from the blades. It then presents the particularities of French regulation of emergence levels and the use of the norm NFS 31-114 during the commissioning acoustic control. It explains the current acoustic model used at the development stage to predict noise emission and curtailment and finally proposes improvements such as considering the topography, the environmental characteristics and the use of uncertainties.

Acoustic models are useful tools to increase understanding of cochlear implant perception. Two particular issues in modelling cochlear implant perception were considered in the present study, which aimed at improving acoustic models. The first included an electricallayer in the model, while the second manipulated synthesis signal parameters. Two parts of the study explored the effects of current decay, compression function and simultaneous stimulation, by including the electrical layer. The SPREAD model, which incorporated this layer, yielded the asymptote in speech intelligibility at seven channels observed in CI listeners. It was shown that the intensity of border channels was deemphasized in relation to more central channels. This was caused by the one-sided effects of current spread from neighbouring channels for the border channels, as opposed to the two-sided effects for the more central channels. It was theorised that more compressive mapping functions would affect spectral cues and consequently speech intelligibility, but speech intelligibility experiments did not confirm this theory. A simultaneous analogue stimulation (SAS) model, which modelled simultaneous stimulation, yielded intelligibility results that were lower than those of the SPREAD model at 16 channels. The SAS model also appeared to introduce more temporal distortion than the SPREAD model. A third part of the study endeavoured to improve correspondence of acoustic model results with cochlear implant listener results by using nine different synthesis signals. The best synthesis signal was noise-band based. The widths of these increased linearly from 0.4 mm at the apical to 8 mm at the basal end. Good correspondence between speech recognition outcomes using this synthesis signal with those of CI listeners was found. AFRIKAANS : Akoesties modelle word algemeen gebruik om die persepsie van inplantingsgebruikers beter te verstaan. Twee benaderings tot die modellering van kogleêre inplantingsgebruikerpersepsies is voorgestel om akoestiese modelle te verbeter. In die eerste benadering is die generiese model verbeter deur die byvoeging van 'n elektriese laag en in die tweede benadering is sinteseseinparameters gemanipuleer om die ooreenkoms met inplantingsgebruikersuitkomste te verbeter. Twee dele van die studie het die effek van stroomverspreiding, samedrukkings-funksie en gelyktydige stimulasie ondersoek deur die insluiting van die elektriese laag. Die SPREAD-model het die asimptoot in spraakherkenning by sewe kanale getoon. Die intensiteit van grenskanale is onderbeklemtoon in verhouding met meer sentrale kanale. Dit is veroorsaak deur die eensydige effekte van stroomverspreiding vir die grenskanale, teenoor die tweesydige effekte wat meer sentrale kanale tipies beïnvloed. Die model het gesuggereer dat meer samedrukkende funksies spektrale inligting sou affekteer, maar spraakherkenningsdata het nie hierdie teorie bevestig nie. Die gelyktydige- analoogstimulasiemodel, wat gelyktydige stimulasie gemodelleer het, het soortgelyke tendense getoon, maar met meer temporale effekte as die SPREAD-model. Die gelyktydige- analoogstimulasiemodel-model se resultate was ook swakker by 16 kanale as die SPREAD-modelresultate. Die derde deel van die studie het gepoog om beter ooreenkoms tussen modeluitkomste en inplantingsgebruikeruitkomste te verkry deur nege verskillende sinteseseine te gebruik. Die beste sintesesein was die ruisband met veranderende wydte; hierdie wydte het verbreed vanaf 0.4 mm by die apeks tot by 8 mm by die basis. 'n Goeie ooreenkoms is verkry tussen modeluitkomste en inplantingsgebruikeruitkomste deur hierdie sintesesein te gebruik. / Thesis (PhD)--University of Pretoria, 2010. / Electrical, Electronic and Computer Engineering / Unrestricted

In this dissertation, three research directions were explored to alleviate two major issues, i.e., the use of incorrect models and training/test condition mismatches, in the modeling frameworks of modern spoken keyword search (KWS) systems. Each of the three research directions, which include (i) data-efficient training processes, (ii) system optimization objectives, and (iii) data augmentation, utilizes different types and amounts of training resources in different ways to ameliorate the two issues of acoustic and language modeling in modern KWS systems. To be more specific, resource-dependent keyword modeling, keyword-boosted sMBR (state-level minimum Bayes risk) training, and multilingual acoustic modeling are proposed and investigated for acoustic modeling in this research. For language modeling, keyword-aware language modeling, discriminative keyword-aware language modeling, and web text augmented language modeling are presented and discussed. The dissertation provides a comprehensive collection of solutions and strategies to the acoustic and language modeling problems in KWS. It also offers insights into the realization of good-performance KWS systems. Experimental results show that the data-efficient training process and data augmentation are the two directions providing the most prominent performance improvement for KWS systems. While modifying system optimization objectives provides smaller yet consistent performance enhancement in KWS systems with different configurations. The effects of the proposed acoustic and language modeling approaches in the three directions are also shown to be additive and can be combined to further improve the overall KWS system performance.

<p>In the modern world, Internet has become a popular place, people with speech hearing disabilities and search engines can't take part of speech content in podcast les. In order to solve the problem partially, the Sphinx decoders such as Sphinx-3, Sphinx-4 can be used to implement a Auto Transcript Generator application, by coupling already existing large acoustic model, language model and a existing dictionary, or by training your own large acoustic model, language model and creating your own dictionary to support continuous speaker independent speech recognition system.</p>

In the modern world, Internet has become a popular place, people with speech hearing disabilities and search engines can't take part of speech content in podcast les. In order to solve the problem partially, the Sphinx decoders such as Sphinx-3, Sphinx-4 can be used to implement a Auto Transcript Generator application, by coupling already existing large acoustic model, language model and a existing dictionary, or by training your own large acoustic model, language model and creating your own dictionary to support continuous speaker independent speech recognition system.

High levels of speech recognition have been obtained with cochlear implant users in quiet conditions. In noisy environments, speech recognition deteriorates considerably, especially in speech-like noise. The aim of this study was to determine what underlies measured speech recognition in cochlear implantees, and furthermore, what underlies perception of speech in noise. Vowel and consonant recognition was determined in ten normal-hearing listeners using acoustic simulations. An acoustic model was developed in order to process vowels and consonants in quiet and noisy conditions; multi-talker babble and speech-like noise were added to the speech segments for the noisy conditions. A total of seven conditions were simulated acoustically; namely for recognition in quiet and as a function of signal-to-noise ratio (0 dB, 20 dB and 40 dB speech-like noise and 0 dB, 20 dB and 40 dB multi-talker babble). An eight- channel SPEAK processor was modelled and used to process the speech segments. A number of biophysical interactions between simulated nerve fibres and the cochlear implant were simulated by including models of these interactions in the acoustic model. Biophysical characteristics that were modelled included dynamic range compression and current spread in the cochlea. Recognition scores deteriorated with increasing noise levels, as expected. Vowel recognition was better than consonant recognition in general. In quiet conditions, the features transmitted most efficiently for recognition of speech segments were duration and F2 for vowels and burst and affrication for consonants. In noisy conditions, listeners mainly depended on the duration of vowels for recognition and the burst of consonants. As the SNR decreased, the number of features used to recognise speech segments also became fewer. This suggests that the addition of noise reduces the number of acoustic features available for recognition. Efforts to improve the transmission of important speech features m cochlear implants should improve recognition of speech in noisy conditions. / Dissertation (MEng (Bio-Engineering))--University of Pretoria, 2008. / Electrical, Electronic and Computer Engineering / unrestricted

People with motor function disorders that cause dysarthric speech find difficulty using state-of- the-art automatic speech recognition (ASR) systems. These systems are developed based on non- dysarthric speech models, which explains the poor performance when used by individuals with dysarthria. Thus, a solution is needed to compensate for the poor performance of these systems. This thesis examines the possibility of quantifying vowels of dysarthric and non-dysarthric speech into codewords regardless of inter-speaker variability and possible to be implemented on limited- processing-capability machines. I show that it is possible to model all possible vowels and vowel- like sounds that a North American speaker can produce if the frequencies of the first and second formants are used to encode these sounds. The proposed solution is aligned with the use of neural networks and hidden Markov models to build an acoustic model in conventional ASR systems. A secondary finding of this study includes the feasibility of reducing the set of ten most common vowels in North American English to eight vowels only. / Graduate / 2021-05-11

A model that can accurately predict speech recognition for cochlear implant (CI) listeners is essential for the optimal fitting of cochlear implants. By implementing a CI acoustic model that mimics CI speech processing, the challenge of predicting speech perception in cochlear implants can be simplified. As a first step in predicting the recognition of speech processed through an acoustic model, vowel perception in severe speech-shaped noise was investigated in the current study. The aim was to determine the acoustic cues that listeners use to recognize vowels in severe noise and make suggestions regarding a vowel perception predictor. It is known that formants play an important role in quiet, while in severe noise the role of formants is still unknown. The relative importance of F1 and F2 is also of interest, since the masking of noise is not always evenly distributed over the vowel spectrum. The problem was addressed by synthesizing vowels consisting of either detailed spectral shape or formant information. F1 and F2 were also suppressed to examine the effect in severe noise. The synthetic stimuli were presented to listeners in quiet and signal-to-noise ratios of 0 dB, -5 dB and -10 dB. Results showed that in severe noise, vowels synthesized according to the whole-spectrum were recognized significantly better than vowels containing only formants. Multidimensional scaling and FITA analysis indicated that formants were still perceived and extracted by the human auditory system in severe noise, especially when the vowel spectrum consisted of the whole spectral shape. Although F1 and F2 vary in importance in listening conditions of quiet and less noisy conditions, the role of the two cues appears to be similar in severe noise. It was suggested that not only the availability formants, but also details of the vowel spectral shape can help to predict vowel recognition in severe noise to a certain degree. / Dissertation (MEng)--University of Pretoria, 2010. / Electrical, Electronic and Computer Engineering / unrestricted

京都大学 / 0048 / 新制・課程博士 / 博士(情報学) / 甲第19849号 / 情博第600号 / 新制||情||104(附属図書館) / 32885 / 京都大学大学院情報学研究科知能情報学専攻 / (主査)教授 河原 達也, 教授 黒橋 禎夫, 教授 鹿島 久嗣 / 学位規則第4条第1項該当 / Doctor of Informatics / Kyoto University / DFAM

Lean premixed fuel-air conditions in large gas turbines are used to improve efficiency and reduce emissions. These conditions give rise to large undamped pressure oscillations at the combustor's natural frequencies which reduce the turbine's longevity and reliability. Active control of the pressure oscillations, called thermoacoustic instabilities, has been sought as passive abatement of these instabilities does not provide adequate damping and is often impractical on a large scale. Phase shift control of the instabilities is perhaps the simplest and most popular technique employed but often does not provide good performance in that controller induced secondary instabilities are generated with increasing loop gain. This thesis investigates the general underlying cause of the secondary instabilities and shows that high average group delay through the frequency region of the instability is the root of the problem. This average group delay is then shown to be due not only the controller itself but can also be associated with other components and inherent characteristics of the control loop such as actuators and time delay, respectively. An "optimum" phase shift controller, consisting of an appropriate shift in phase and a low order, wide bandwidth bandpass filter, is developed for a Rijke tube combustor and shown to closely match the response of an LQG controller designed only for system stabilization. Both the optimal phase shifter and the LQG controller are developed based on a modified model of the thermoacoustic loop which takes into account the change in density of the combustion reactants at the flame location. Additionally, the system model is coupled with a model of the control loop and then validated by comparison of simulated results to experimental results using nearly identical controllers. / Master of Science