Accuracy Of Automatic Recognition Research Articles

Automatic Recognition of Auditory Brainstem Response Waveforms Using a Deep Learning-Based Framework.

Recognition of auditory brainstem response (ABR) waveforms may be challenging, particularly for older individuals or those with hearing loss. This study aimed to investigate deep learning frameworks to improve the automatic recognition of ABR waveforms in participants with varying ages and hearing levels. The research used a descriptive study design to collect and analyze pure tone audiometry and ABR data from 100 participants. The research was conducted at a tertiary academic medical center, specifically at the Clinical Audiology Center of Tsinghua Chang Gung Hospital (Beijing, China). Data from 100 participants were collected and categorized into four groups based on age and hearing level. Features from both time-domain and frequency-domain ABR signals were extracted and combined with demographic factors, such as age, sex, pure-tone thresholds, stimulus intensity, and original signal sequences to generate feature vectors. An enhanced Wide&Deep model was utilized, incorporating the Light-multi-layer perceptron (MLP) model to train the recognition of ABR waveforms. The recognition accuracy (ACC) of each model was calculated for the overall data set and each group. The ACC rates of the Light-MLP model were 97.8%, 97.2%, 93.8%, and 92.0% for Groups 1 to 4, respectively, with a weighted average ACC rate of 95.4%. For the Wide&Deep model, the ACC rates were 93.4%, 90.8%, 92.0%, and 88.3% for Groups 1 to 4, respectively, with a weighted average ACC rate of 91.0%. Both the Light-MLP model and the Wide&Deep model demonstrated excellent ACC in automatic recognition of ABR waveforms across participants with diverse ages and hearing levels. While the Wide&Deep model's performance was slightly poorer than that of the Light-MLP model, particularly due to the limited sample size, it is anticipated that with an expanded data set, the performance of Wide&Deep model may be further improved.

A segmentally informed solution to automatic accent classification and its advantages to forensic applications

Traditionally, work in automatic accent recognition has followed a similar research trajectory to that of language identification, dialect identification and automatic speaker recognition. The same acoustic modelling approaches that have been implemented in speaker recognition (such as GMM-UBM and i-vector-based systems) have also been applied to automatic accent recognition. These approaches form models of speakers’ accents by taking acoustic features from right across the speech signal without knowledge of its phonetic content. Particularly for accent recognition, however, phonetic information is expected to add substantial value to the task. The current work presents an alternative modelling approach to automatic accent recognition, which forms models of speakers’ pronunciation systems using segmental information. This article claims that such an approach to the problem makes for a more explainable method and therefore is a more appropriate method to deploy in settings where it is important to be able to communicate methods, such as forensic applications. We discuss the issue of explainability and show how the system operates on a large 700-speaker dataset of non-native English conversational telephone recordings.

Computational modelling of segmental and prosodic levels of analysis for capturing variation across Arabic dialects

Dialect variation spans different linguistic levels of analysis. Two examples include the typical phonetic realisations produced and the typical range of intonational choices made by individuals belonging to a given dialect group. Taking the modeling principles of a specific automatic accent recognition system, the work here characterises and observes the variation that exists within these two levels of analysis among eight Arabic dialects. Using a method that has previously shown promising performance on English accent varieties, we first model the segmental level of analysis from recordings of Arabic speakers to capture the variation in the phonetic realisations of the vowels and consonants. In doing so, we show how powerful this model can be in distinguishing between Arabic dialects. This paper then shows how this modeling approach can be adapted to instead characterise prosodic variation among these same dialects from the same speech recordings. This allows us to inspect the relative power of the segmental and prosodic levels of analysis in separating the Arabic dialects. This work opens up the possibility of using these modeling frameworks to study the extent and nature of phonetic and prosodic variation across speech corpora.

A Novel Method Based on Deep Learning, GIS and Geomatics Software for Building a 3D City Model from VHR Satellite Stereo Imagery

The aim of the paper is to identify a suitable method for the construction of a 3D city model from stereo satellite imagery. In order to reach this goal, it is necessary to build a workflow consisting of three main steps: (1) Increasing the geometric resolution of the color images through the use of pan-sharpening techniques, (2) identification of the buildings’ footprint through deep-learning techniques and, finally, (3) building an algorithm in GIS (Geographic Information System) for the extraction of the elevation of buildings. The developed method was applied to stereo imagery acquired by WorldView-2 (WV-2), a commercial Earth-observation satellite. The comparison of the different pan-sharpening techniques showed that the Gram–Schmidt method provided better-quality color images than the other techniques examined; this result was deduced from both the visual analysis of the orthophotos and the analysis of quality indices (RMSE, RASE and ERGAS). Subsequently, a deep-learning technique was applied for pan sharpening an image in order to extract the footprint of buildings. Performance indices (precision, recall, overall accuracy and the F1measure) showed an elevated accuracy in automatic recognition of the buildings. Finally, starting from the Digital Surface Model (DSM) generated by satellite imagery, an algorithm built in the GIS environment allowed the extraction of the building height from the elevation model. In this way, it was possible to build a 3D city model where the buildings are represented as prismatic solids with flat roofs, in a fast and precise way.

Research on Basketball Goal Recognition Based on Image Processing and Improved Algorithm

This paper studies the basketball goal recognition method based on image processing and improved algorithm to improve the accuracy of automatic recognition of basketball goal. The infrared spectrum image acquisition system is used to collect the basketball goal image. After the image is denoised by using the adaptive filtering algorithm, the wavelet analysis method is used to extract the features of basketball goal signal, which are input into the optimized deformable convolution neural network. Through the weighted sum of the values of each sampling point and the corresponding position authority of the block convolution core, the results are output as convolution operation. Combined with the depth feature of the same dimension, the full connection feature of the candidate target area is obtained to realize the basketball goal recognition. The experimental results show the following: the method can effectively identify basketball goals and the recognition error rate is low; the average accuracy of the automatic recognition results of basketball goals is as high as 98.4%; under the influence of different degrees of noise, the method is less affected by noise and has strong anti-interference ability.

Automated Mapping of Ms 7.0 Jiuzhaigou Earthquake (China) Post-Disaster Landslides Based on High-Resolution UAV Imagery

The Ms 7.0 Jiuzhaigou earthquake that occurred on 8 August 2017 triggered hundreds of landslides in the Jiuzhaigou valley scenic and historic-interest area in Sichuan, China, causing heavy casualties and serious property losses. Quick and accurate mapping of post-disaster landslide distribution is of paramount importance for earthquake emergency rescue and the analysis of post-seismic landslides distribution characteristics. The automatic identification of landslides is mostly based on medium- and low-resolution satellite-borne optical remote-sensing imageries, and the high-accuracy interpretation of earthquake-triggered landslides still relies on time-consuming manual interpretation. This paper describes a methodology based on the use of 1 m high-resolution unmanned aerial vehicle (UAV) imagery acquired after the earthquake, and proposes a support vector machine (SVM) classification method combining the roads and villages mask from pre-seismic remote sensing imagery to accurately and automatically map the landslide inventory. Compared with the results of manual visual interpretation, the automatic recognition accuracy could reach 99.89%, and the Kappa coefficient was higher than 0.9, suggesting that the proposed method and 1 m high-resolution UAV imagery greatly improved the mapping accuracy of the landslide area. We also analyzed the spatial-distribution characteristics of earthquake-triggered landslides with the influenced factors of altitude, slope gradient, slope aspect, and the nearest faults, which provided important support for the further study of post-disaster landslide distribution characteristics, susceptibility prediction, and risk assessment.

NU-LiteNet: Mobile Landmark Recognition using Convolutional Neural Networks

The growth of high-performance mobile devices has resulted in more research into on-device image recognition. The research problems have been the latency and accuracy of automatic recognition, which remain as obstacles to its real-world usage. Although the recently developed deep neural networks can achieve accuracy comparable to that of a human user, some of them are still too slow. This paper describes the development of the architecture of a new convolutional neural network model, NU-LiteNet. For this, SqueezeNet was developed to reduce the model size to a degree suitable for smartphones. The model size of NU-LiteNet was therefore 2.6 times smaller than that of SqueezeNet. The model outperformed other Convolutional Neural Network (CNN) models for mobile devices (eg. SqueezeNet and MobileNet) with an accuracy of 81.15% and 69.58% on Singapore and Paris landmark datasets respectively. The shortest execution time of 0.7 seconds per image was recorded with NU-LiteNet on mobile phones.

Automatic sociophonetics: Exploring corpora with a forensic accent recognition system.

This paper demonstrates how the Y-ACCDIST system, the York ACCDIST-based automatic accent recognition system [Brown (2015). Proceedings of the International Congress of Phonetic Sciences, Glasgow, UK], can be used to inspect sociophonetic corpora as a preliminary "screening" tool. Although Y-ACCDIST's intended application is to assist with forensic casework, the system can also be exploited in sociophonetic research to begin unpacking variation. Using a subset of the PEBL (Panjabi-English in Bradford and Leicester) corpus, the outputs of Y-ACCDIST are explored, which, it is argued, efficiently and objectively assess speaker similarities across different linguistic varieties. The ways these outputs corroborate with a phonetic analysis of the data are also discovered. First, Y-ACCDIST is used to classify speakers from the corpus based on language background and region. A Y-ACCDIST cluster analysis is then implemented, which groups speakers in ways consistent with more localised networks, providing a means of identifying potential communities of practice. Additionally, the results of a Y-ACCDIST feature selection task that indicates which specific phonemes are most valuable in distinguishing between speaker groups are presented. How Y-ACCDIST outputs can be used to reinforce more traditional sociophonetic analyses and support qualitative interpretations of the data is demonstrated.

Distant-talking accent recognition by combining GMM and DNN

Recently, automatic accent recognition has been paid more and more attentions. However, there are few researches focusing on accent recognition in distant-talking environment which is very important for improving distant-talking speech recognition performance with non-native accents. In this paper, we apply Gaussian Mixture Models (GMM) and Deep Neural Network (DNN) to identify the speaker accent in reverberant environments. The combination of likelihood with these two approaches is also proposed. In reverberant environment, the accent recognition rate was improved from 90.7 % with GMM to 93.0 % with DNN. The combination of GMM and DNN achieved recognition rate of 97.5 %, which outperformed than the individual GMM and DNN because the complementation of GMM and DNN. The relative error reduction is 73.1 % than the GMM-based method and 64.3 % than the DNN-based method, respectively.

Human and computer recognition of regional accents and ethnic groups from British English speech

The paralinguistic information in a speech signal includes clues to the geographical and social background of the speaker. This paper is concerned with automatic extraction of this information from a short segment of speech. A state-of-the-art language identification (LID) system is applied to the problems of regional accent recognition for British English, and ethnic group recognition within a particular accent. We compare the results with human performance and, for accent recognition, the ‘text dependent’ ACCDIST accent recognition measure. For the 14 regional accents of British English in the ABI-1 corpus (good quality read speech), our LID system achieves a recognition accuracy of 89.6%, compared with 95.18% for our best ACCDIST-based system and 58.24% for human listeners. The “Voices across Birmingham” corpus contains significant amounts of telephone conversational speech for the two largest ethnic groups in the city of Birmingham (UK), namely the ‘Asian’ and ‘White’ communities. Our LID system distinguishes between these two groups with an accuracy of 96.51% compared with 90.24% for human listeners. Although direct comparison is difficult, it seems that our LID system performs much better on the standard 12 class NIST 2003 Language Recognition Evaluation task or the two class ethnic group recognition task than on the 14 class regional accent recognition task. We conclude that automatic accent recognition is a challenging task for speech technology, and speculate that the use of natural conversational speech may be advantageous for these types of paralinguistic task.

Speech enhancement for noise‐robust speech recognition.

A modified phase opponency (MPO) model was combined with the aperiodic/periodic/pitch detector (APP) by Deshmukh et al. [JASA, 121, 3886–3898 (2006)] for speech enhancement. The MPO‐APP passes the speech‐dominated regions as is but greatly attenuates the noise‐dominated regions. Thus, the noise in nonspeech regions and in the regions between the formants are significantly reduced. Unlike many other techniques, the MPO‐APP does not introduce musical noise. However, given that the noise in formant regions is passed along with the speech signal, the enhanced signal sometimes has a shadow effect. To get rid of this shadow effect, we recently developed a preprocessor for the MPO‐APP. First, the signal to noise ratio of the signal is estimated. Second, we use an adaptive generalized spectral subtraction technique to reduce the noise in the speech dominant regions. The results show a significant improvement in the quality of the enhanced signal and, in addition, increased accuracy in automatic recognition of the enhanced signals over that obtained from the original signals or the signals enhanced without the preprocessor. [Research supported by NSF Grant IIS0703859.]