Log Mel-filter Bank Research Articles

Derin Öğrenme İle Türkçe Müziklerden Müzik Türü Sınıflandırması

In this study, an architecture called Convolutional Long Short-term memory deep neural network (CLDNN) based on deep learning, which has not been used before in this field, is used for music genre classification. In addition, a new Turkish Music Database consisting of 200 music belonging to various music genres has been created. The classification performance of the proposed architecture and commonly used machine learning methods has been evaluated on this database. In addition, new features are obtained by using Convolutional Neural Network (CNN), which is the first part of this architecture. Both Mel Frequency Cepstrum Coefficients (MFCC) and log mel filterbank energies were used as input to the Convolutional Neural Network to obtain these new features. In addition to these features, many standard features have been obtained by using various toolboxes. The most successful classification results for all methods are achieved when standard features are used together with new features. The best results among the compared classifiers were achieved with 99.5% by using the remaining part of the proposed architecture, Long Short Term Memory (LSTM), together with the Deep Neural Network (DNN) consisting of fully connected layers.

Human activity classification based on sound recognition and residual convolutional neural network

Human activity recognition is crucial for a better understanding of workers in construction sites and people in the built environment. Previous studies have been proposed various ways in which sensing and machine learning techniques can be utilized to collect human activity data automatically. Sound recognition has the potential to be utilized in ways that complement the limitations of the previous methods because sound signals are easy to propagate in indoor environments where many physical obstacles exist, and this method can simultaneously recognize not only sounds from human activities but also sounds from related objects. Therefore, this study develops a sound recognition-based human activity classification model using a residual neural network. A sound data is collected based on ten classes representing people's daily activities in the indoor environment. Then, the features of the sound data were extracted using the Log Mel-filter bank energies method, and a residual neural network model with 34 convolutional layers was trained using the data. The results showed the following: the accuracy of the model was 87.6%, and the Precision score for each class ranged from 76.8% to 92.6%, the Recall scores ranged from 75.8% to 98.6%, and the F1-score ranged from 78.6% to 93.7%. The contribution of this study is to demonstrate that sound recognition can classify people's indoor activities successfully, but this study leaves the limitation that it is based on a monophonic method that only one activity can be classified at a time.

Baby Cry Detection in Domestic Environment Using Deep Learning

Automatic detection of a baby cry in audio signals is an essential step in applications such as remote baby monitoring. It is also important for researchers, who study the relation between baby cry patterns and various health or developmental parameters. In this paper, we propose two machine-learning algorithms for automatic detection of baby cry in audio recordings. The first algorithm is a low-complexity logistic regression classifier, used as a reference. To train this classifier, we extract features such as Mel-frequency cepstrum coefficients, pitch and formants from the recordings. The second algorithm uses a dedicated convolutional neural network (CNN), operating on log Mel-filter bank representation of the recordings. Performance evaluation of the algorithms is carried out using an annotated database containing recordings of babies (0-6 months old) in domestic environments. In addition to baby cry, these recordings contain various types of domestic sounds, such as parents talking and door opening. The CNN classifier is shown to yield considerably better results compared to the logistic regression classifier, demonstrating the power of deep learning when applied to audio processing.

Drum sound onset detection based on class separation using deep neural network

This study reports an onset detection method for the snare drum sounds for music signals constructed with plural instruments. We propose a method which detects an onset with a class separation using a deep neural network (DNN). To train the network, frames which include the waveform of the snare drum were extracted from music signal, and the log mel-filter bank channel outputs were calculated for each frame. The DNN was trained by using these feature parameters. The inputs were the calculated filter bank outputs, and the teaching signals were binary information whether to include the waveform of a snare drum or not. To detect onsets, candidate frames of the onset time were found based on changing in the power of the target music signal, and feature values were extracted from the candidate frames. Obtained values were used as the input data for the trained DNN, and detection of whether or not to include a waveform of a snare drum can be achieved. Experiments were conducted for 52 popular songs included in the music database, and an average F-measure was 0.73.

Environmentally robust ASR front-end for deep neural network acoustic models

This paper examines the individual and combined impacts of various front-end approaches on the performance of deep neural network (DNN) based speech recognition systems in distant talking situations, where acoustic environmental distortion degrades the recognition performance. Training of a DNN-based acoustic model consists of generation of state alignments followed by learning the network parameters. This paper first shows that the network parameters are more sensitive to the speech quality than the alignments and thus this stage requires improvement. Then, various front-end robustness approaches to addressing this problem are categorised based on functionality. The degree to which each class of approaches impacts the performance of DNN-based acoustic models is examined experimentally. Based on the results, a front-end processing pipeline is proposed for efficiently combining different classes of approaches. Using this front-end, the combined effects of different classes of approaches are further evaluated in a single distant microphone-based meeting transcription task with both speaker independent (SI) and speaker adaptive training (SAT) set-ups. By combining multiple speech enhancement results, multiple types of features, and feature transformation, the front-end shows relative performance gains of 7.24% and 9.83% in the SI and SAT scenarios, respectively, over competitive DNN-based systems using log mel-filter bank features.

Feature Mapping of Multiple Beamformed Sources for Robust Overlapping Speech Recognition Using a Microphone Array

This paper introduces a nonlinear vector-based feature mapping approach to extract robust features for automatic speech recognition (ASR) of overlapping speech using a microphone array. We explore different configurations and additional sources of information to improve the effectiveness of the feature mapping. First, we investigate the full-vector based mapping of different sources in a log mel-filterbank energy (log MFBE) domain, and demonstrate that retraining the acoustic model using the generated training data can help improve the recognition performance. Then we investigate the feature mapping between different domains. Finally in order to improve the qualities of the mapping inputs we propose a nonlinear mapping of the features from multiple beamformed sources, which are directed at the target and interfering speakers, respectively. We demonstrate the effectiveness of the proposed approach through extensive evaluations on the MONC corpus, which includes non-overlapping single speaker and overlapping multi-speaker conditions.

Feature enhancement by deep LSTM networks for ASR in reverberant multisource environments

This article investigates speech feature enhancement based on deep bidirectional recurrent neural networks. The Long Short-Term Memory (LSTM) architecture is used to exploit a self-learnt amount of temporal context in learning the correspondences of noisy and reverberant with undistorted speech features. The resulting networks are applied to feature enhancement in the context of the 2013 2nd Computational Hearing in Multisource Environments (CHiME) Challenge track 2 task, which consists of the Wall Street Journal (WSJ-0) corpus distorted by highly non-stationary, convolutive noise. In extensive test runs, different feature front-ends, network training targets, and network topologies are evaluated in terms of frame-wise regression error and speech recognition performance. Furthermore, we consider gradually refined speech recognition back-ends from baseline ‘out-of-the-box’ clean models to discriminatively trained multi-condition models adapted to the enhanced features. In the result, deep bidirectional LSTM networks processing log Mel filterbank outputs deliver best results with clean models, reaching down to 42% word error rate (WER) at signal-to-noise ratios ranging from −6 to 9dB (multi-condition CHiME Challenge baseline: 55% WER). Discriminative training of the back-end using LSTM enhanced features is shown to further decrease WER to 22%. To our knowledge, this is the best result reported for the 2nd CHiME Challenge WSJ-0 task yet.