Identifying musical instruments in polyphonic music recordings is a challenging but important problem in the field of music information retrieval. It enables music search by instrument, helps recognize musical genres, or can make music transcription easier and more accurate. In this paper, we present a convolutional neural network framework for predominant instrument recognition in real-world polyphonic music. We train our network from fixed-length music excerpts with a single-labeled predominant instrument and estimate an arbitrary number of predominant instruments from an audio signal with a variable length. To obtain the audio-excerpt-wise result, we aggregate multiple outputs from sliding windows over the test audio. In doing so, we investigated two different aggregation methods: one takes the class-wise average followed by normalization, and the other perform temporally local class-wise max-pooling on the output probability prior to averaging and normalization steps to minimize the effect of averaging process suppresses the activation of sporadically appearing instruments. In addition, we conducted extensive experiments on several important factors that affect the performance, including analysis window size, identification threshold, and activation functions for neural networks to find the optimal set of parameters. Our analysis on the instrument-wise performance found that the onset type is a critical factor for recall and precision of each instrument. Using a dataset of 10k audio excerpts from 11 instruments for evaluation, we found that convolutional neural networks are more robust than conventional methods that exploit spectral features and source separation with support vector machines. Experimental results showed that the proposed convolutional network architecture obtained an F1 measure of 0.619 for micro and 0.513 for macro, respectively, achieving 23.1% and 18.8% in performance improvement compared with the state-of-the-art algorithm.

Breathing sounds during sleep are altered and characterized by various acoustic specificities in patients with sleep disordered breathing (SDB). This study aimed to identify acoustic biomarkers indicative of the severity of SDB by analyzing the breathing sounds collected from a large number of subjects during entire overnight sleep. The participants were patients who presented at a sleep center with snoring or cessation of breathing during sleep. They were subjected to full-night polysomnography (PSG) during which the breathing sound was recorded using a microphone. Then, audio features were extracted and a group of features differing significantly between different SDB severity groups was selected as a potential acoustic biomarker. To assess the validity of the acoustic biomarker, classification tasks were performed using several machine learning techniques. Based on the apnea–hypopnea index of the subjects, four-group classification and binary classification were performed. Using tenfold cross validation, we achieved an accuracy of 88.3% in the four-group classification and an accuracy of 92.5% in the binary classification. Experimental evaluation demonstrated that the models trained on the proposed acoustic biomarkers can be used to estimate the severity of SDB. Acoustic biomarkers may be useful to accurately predict the severity of SDB based on the patient’s breathing sounds during sleep, without conducting attended full-night PSG. This study implies that any device with a microphone, such as a smartphone, could be potentially utilized outside specialized facilities as a screening tool for detecting SDB.

Artificial reverberation (AR) models play a central role in various audio applications. Therefore, estimating the AR model parameters (ARPs) of a reference reverberation is a crucial task. Although a few recent deep-learning-based approaches have shown promising performance, their non-end-to-end training scheme prevents them from fully exploiting the potential of deep neural networks. This motivates the introduction of differentiable artificial reverberation (DAR) models, allowing loss gradients to be back-propagated end-to-end. However, implementing the AR models with their difference equations “as is” in the deep learning framework severely bottlenecks the training speed when executed with a parallel processor like GPU due to their infinite impulse response (IIR) components. We tackle this problem by replacing the IIR filters with finite impulse response (FIR) approximations with the frequency-sampling method. Using this technique, we implement three DAR models—differentiable Filtered Velvet Noise (FVN), Advanced Filtered Velvet Noise (AFVN), and Delay Network (DN). For each AR model, we train its ARP estimation networks for analysis-synthesis (RIR-to-ARP) and blind estimation (reverberant-speech-to-ARP) task in an end-to-end manner with its DAR model counterpart. Experiment results show that the proposed method achieves consistent performance improvement over the non-end-to-end approaches in both objective metrics and subjective listening test results.

Recent deep learning approaches for melody harmonization have achieved remarkable performance by overcoming the uneven chord distributions of music data. However, most of these approaches have not attempted to capture an original melodic structure and generate structured chord sequences with appropriate rhythms. Hence, we use a Transformer-based architecture that directly maps lower-level melody notes into a semantic higher-level chord sequence. In particular, we encode the binary piano roll of a melody into a note-based representation. Furthermore, we address the flexible generation of various chords with Transformer expanded with a VAE framework. We propose three Transformer-based melody harmonization models: 1) the standard Transformer-based model for the neural translation of a melody to chords (STHarm), 2) the variational Transformer model for learning the global representation of complete music (VTHarm), and 3) the regularized variational Transformer model for the controllable generation of chords (rVTHarm). Experimental results demonstrate that the proposed models generate more structured, diverse chord sequences than LSTM-based models.

Quantitative evaluation of piano performance is of interests in many fields, including music education and computational performance rendering. Previous studies utilized features extracted from audio or musical instrument digital interface (MIDI) files but did not address the difference between hands (DBH), which might be an important aspect of high-quality performance. Therefore, we investigated DBH as an important factor determining performance proficiency. To this end, 34 experts and 34 amateurs were recruited to play two excerpts on a Yamaha Disklavier. Each performance was recorded in MIDI, and handcrafted features were extracted separately for the right hand (RH) and left hand (LH). These were conventional MIDI features representing temporal and dynamic attributes of each note and computed as absolute values (e. g., MIDI velocity) or ratios between performance and corresponding scores (e. g., ratio of duration or inter-onset interval (IOI)). These note-based features were rearranged into additional features representing DBH by simple subtraction between features of both hands. Statistical analyses showed that DBH was more significant in experts than in amateurs across features. Regarding temporal features, experts pressed keys longer and faster with the RH than did amateurs. Regarding dynamic features, RH exhibited both greater values and a smoother change along melodic intonations in experts that in amateurs. Further experiments using principal component analysis (PCA) and support vector machine (SVM) verified that hand-difference features can successfully differentiate experts from amateurs according to performance proficiency. Moreover, existing note-based raw feature values (Basic features) and DBH features were tested repeatedly via 10-fold cross-validation, suggesting that adding DBH features to Basic features improved F1 scores to 93.6% (by 3.5%) over Basic features. Our results suggest that differently controlling both hands simultaneously is an important skill for pianists; therefore, DBH features should be considered in the quantitative evaluation of piano performance.