Animal vocalization denoising is a task similar to human speech enhancement, which is relatively well-studied. In contrast to the latter, it comprises a higher diversity of sound production mechanisms and recording environments, and this higher diversity is a challenge for existing models. Adding to the challenge and in contrast to speech, we lack large and diverse datasets comprising clean vocalizations. As a solution we use as training data pseudo-clean targets, i.e. pre-denoised vocalizations, and segments of background noise without a vocalization. We propose a train set derived from bioacoustics datasets and repositories representing diverse species, acoustic environments, geographic regions. Additionally, we introduce a non-overlapping benchmark set comprising clean vocalizations from different taxa and noise samples. We show that that denoising models (demucs, CleanUNet) trained on pseudo-clean targets obtained with speech enhancement models achieve competitive results on the benchmarking set. We publish data, code, libraries, and demos at https://mariusmiron.com/research/biodenoising.

The use of machine learning (ML) based techniques has become increasingly popular in the field of bioacoustics over the last years. Fundamental requirement for the successful application of ML based techniques are curated, agreed upon, high-quality datasets and benchmark tasks to be learned on a given dataset. However, the field of bioacoustics so far lacks such public benchmarks which cover multiple tasks and species to measure the performance of ML techniques in a controlled and standardized way and that allows for benchmarking newly proposed techniques to existing ones. Here, we propose BEANS (the BEnchmark of ANimal Sounds), a collection of bioacoustics tasks and public datasets, specifically designed to measure the performance of machine learning algorithms in the field of bioacoustics. The benchmark consists of two common tasks in bioacoustics: classification and detection. It includes 12 datasets covering various species, including birds, land and marine mammals, anurans, and insects. In addition to the datasets, we also present the performance of a set of standard ML methods as the baseline for task performance. The benchmark and baseline code is made publicly available1in the hope of establishing a new standard dataset for ML-based bioacoustic research.

In this paper, we propose a novel neural network model called KaraSinger for a less-studied singing voice synthesis (SVS) task named score-free SVS, in which the prosody and melody are spontaneously decided by machine. KaraSinger comprises a vector-quantized variational autoencoder (VQ-VAE) that compresses the Mel-spectrograms of singing audio to sequences of discrete codes, and a language model (LM) that learns to predict the discrete codes given the corresponding lyrics. For the VQ-VAE part, we employ a Connectionist Temporal Classification (CTC) loss to encourage the discrete codes to carry phoneme-related information. For the LM part, we use location-sensitive attention for learning a robust alignment between the input phoneme sequence and the output discrete code. We keep the architecture of both the VQ-VAE and LM light-weight for fast training and inference speed. We validate the effectiveness of the proposed design choices using a proprietary collection of 550 English pop songs sung by multiple amateur singers. The result of a listening test shows that KaraSinger achieves high scores in intelligibility, musicality, and the overall quality.

To apply neural sequence models such as the Transformers to music generation tasks, one has to represent a piece of music by a sequence of tokens drawn from a finite set of pre-defined vocabulary. Such a vocabulary usually involves tokens of various types. For example, to describe a musical note, one needs separate tokens to indicate the note’s pitch, duration, velocity (dynamics), and placement (onset time) along the time grid. While different types of tokens may possess different properties, existing models usually treat them equally, in the same way as modeling words in natural languages. In this paper, we present a conceptually different approach that explicitly takes into account the type of the tokens, such as note types and metric types. And, we propose a new Transformer decoder architecture that uses different feed-forward heads to model tokens of different types. With an expansion-compression trick, we convert a piece of music to a sequence of compound words by grouping neighboring tokens, greatly reducing the length of the token sequences. We show that the resulting model can be viewed as a learner over dynamic directed hypergraphs. And, we employ it to learn to compose expressive Pop piano music of full-song length (involving up to 10K individual tokens per song), both conditionally and unconditionally. Our experiment shows that, compared to state-of-the-art models, the proposed model converges 5 to 10 times faster at training (i.e., within a day on a single GPU with 11 GB memory), and with comparable quality in the generated music

In a recent paper, we have presented a generative adversarial network (GAN)-based model for unconditional generation of the mel-spectrograms of singing voices. As the generator of the model is designed to take a variable-length sequence of noise vectors as input, it can generate mel-spectrograms of variable length. However, our previous listening test shows that the quality of the generated audio leaves room for improvement. The present paper extends and expands that previous work in the following aspects. First, we employ a hierarchical architecture in the generator to induce some structure in the temporal dimension. Second, we introduce a cycle regularization mechanism to the generator to avoid mode collapse. Third, we evaluate the performance of the new model not only for generating singing voices, but also for generating speech voices. Evaluation result shows that new model outperforms the prior one both objectively and subjectively. We also employ the model to unconditionally generate sequences of piano and violin music and find the result promising. Audio examples, as well as the code for implementing our model, will be publicly available online upon paper publication.

Stacked dilated convolutions used in Wavenet have been shown effective for generating high-quality audios. By replacing pooling/striding with dilation in convolution layers, they can preserve high-resolution information and still reach distant locations. Producing high-resolution predictions is also crucial in music source separation, whose goal is to separate different sound sources while maintain the quality of the separated sounds. Therefore, in this paper, we use stacked dilated convolutions as the backbone for music source separation. Although stacked dilated convolutions can reach wider context than standard convolutions do, their effective receptive fields are still fixed and might not be wide enough for complex music audio signals. To reach even further information at remote locations, we propose to combine a dilated convolution with a modified GRU called Dilated GRU to form a block. A Dilated GRU receives information from k-step before instead of the previous step for a fixed k. This modification allows a GRU unit to reach a location with fewer recurrent steps and run faster because it can execute in parallel partially. We show that the proposed model with a stack of such blocks performs equally well or better than the state-of-the-art for separating both vocals and accompaniment.

Music videos are one of the most popular types of video streaming services, and instrument playing is among the most common scenes in such videos. In order to understand the instrument-playing scenes in the videos, it is important to know what instruments are played, when they are played, and where the playing actions occur in the scene. While audio-based recognition of instruments has been widely studied, the visual aspect of music instrument playing remains largely unaddressed in the literature. One of the main obstacles is the difficulty in collecting annotated data of the action locations for training-based methods. To address this issue, we propose a weakly supervised framework to find when and where the instruments are played in the videos. We propose using two auxiliary models: 1) a sound model and 2) an object model to provide supervision for training the instrument-playing action model. The sound model provides temporal supervisions, while the object model provides spatial supervisions. They together can simultaneously provide temporal and spatial supervisions. The resulting model only needs to analyze the visual part of a music video to deduce which, when, and where instruments are played. We found that the proposed method significantly improves localization accuracy. We evaluate the result of the proposed method temporally and spatially on a small dataset (a total of 5400 frames) that we manually annotated.

Convolutional neural networks with skip connections have shown good performance in music source separation. In this work, we propose a denoising Auto-encoder with Recurrent skip Connections (ARC). We use 1D convolution along the temporal axis of the time-frequency feature map in all layers of the fully-convolutional network. The use of 1D convolution makes it possible to apply recurrent layers to the intermediate outputs of the convolution layers. In addition, we also propose an enhancement network and a residual regression method to further improve the separation result. The recurrent skip connections, the enhancement module, and the residual regression all improve the separation quality. The ARC model with residual regression achieves 5.74 siganl-to-distoration ratio (SDR) in vocals with MUSDB (used in SiSEC 2018). We also evaluate the ARC model alone on the older dataset DSD100 (used in SiSEC 2016) and it achieves 5.91 SDR in vocals.

Audio event detection aims at discovering the elements inside an audio clip. In addition to labeling the clips with the audio events, we want to find out the temporal locations of these events. However, creating clearly annotated training data can be time-consuming. Therefore, we provide a model based on convolutional neural networks that relies only on weakly-supervised data for training. These data can be directly obtained from online platforms, such as Freesound, with the clip-level labels assigned by the uploaders. The structure of our model is extended to a fully convolutional networks, and an event-specific Gaussian filter layer is designed to advance its learning ability. Besides, this model is able to detect frame-level information, e.g., the temporal position of sounds, even when it is trained merely with clip-level labels.

Being able to predict whether a song can be a hit has important applications in the music industry. Although it is true that the popularity of a song can be greatly affected by external factors such as social and commercial influences, to which degree audio features computed from musical signals (whom we regard as internal factors) can predict song popularity is an interesting research question on its own. Motivated by the recent success of deep learning techniques, we attempt to extend previous work on hit song prediction by jointly learning the audio features and prediction models using deep learning. Specifically, we experiment with a convolutional neural network model that takes the primitive mel-spectrogram as the input for feature learning, a more advanced JYnet model that uses an external song dataset for supervised pre-training and auto-tagging, and the combination of these two models. We also consider the inception model to characterize audio information in different scales. Our experiments suggest that deep structures are indeed more accurate than shallow structures in predicting the popularity of either Chinese or Western Pop songs in Taiwan. We also use the tags predicted by JYnet to gain insights into the result of different models.

In music auto-tagging, people develop models to automatically label a music clip with attributes such as instruments, styles or acoustic properties. Many of these tags are actually descriptors of local events in a music clip, rather than a holistic description of the whole clip. Localizing such tags in time can potentially innovate the way people retrieve and interact with music, but little work has been done to date due to the scarcity of labeled data with granularity specific enough to the frame level. Most labeled data for training a learning-based model for music auto-tagging are in the clip level, providing no cues when and how long these attributes appear in a music clip. To bridge this gap, we propose in this paper a convolutional neural network (CNN) architecture that is able to make accurate frame-level predictions of tags in unseen music clips by using only clip-level annotations in the training phase. Our approach is motivated by recent advances in computer vision for localizing visual objects, but we propose new designs of the CNN architecture to account for the temporal information of music and the variable duration of such local tags in time. We report extensive experiments to gain insights into the problem of event localization in music, and validate through experiments the effectiveness of the proposed approach. In addition to quantitative evaluations, we also present qualitative analyses showing the model can indeed learn certain characteristics of music tags.

Recent years have witnessed a growing interest in modeling user behaviors in multimedia research, emphasizing the need to consider human factors such as preference, activity, and emotion in system development and evaluation. Following this research line, we present in this paper the LiveJournal two-million post (LJ2M) dataset to foster research on user-centered music information retrieval. The new dataset is characterized by the great diversity of real-life listening contexts where people and music interact. It contains blog articles from the social blogging website LiveJournal, along with tags self-reporting a user's emotional state while posting and the musical track that the user considered as the best match for the post. More importantly, the data are contributed by users spontaneously in their daily lives, instead of being collected in a controlled environment. Therefore, it offers new opportunities to understand the interrelationship among the personal, situational, and musical factors of music listening. As an example application, we present research investigating the interaction between the affective context of the listener and the affective content of music, using audio-based music emotion recognition techniques and a psycholinguistic tool. The study offers insights into the role of music in mood regulation and demonstrates how LJ2M can contribute to studies on real-world music listening behavior.

There has been an increasing attention on learning feature representations from the complex, high-dimensional audio data applied in various music information retrieval (MIR) problems. Unsupervised feature learning techniques, such as sparse coding and deep belief networks have been utilized to represent music information as a term-document structure comprising of elementary audio codewords. Despite the widespread use of such bag-of-frames (BoF) model, few attempts have been made to systematically compare different component settings. Moreover, whether techniques developed in the text retrieval community are applicable to audio codewords is poorly understood. To further our understanding of the BoF model, we present in this paper a comprehensive evaluation that compares a large number of BoF variants on three different MIR tasks, by considering different ways of low-level feature representation, codebook construction, codeword assignment, segment-level and song-level feature pooling, tf-idf term weighting, power normalization, and dimension reduction. Our evaluations lead to the following findings: 1) modeling music information by two levels of abstraction improves the result for difficult tasks such as predominant instrument recognition, 2) tf-idf weighting and power normalization improve system performance in general, 3) topic modeling methods such as latent Dirichlet allocation does not work for audio codewords.

This paper proposes a context-aware approach that recommends music to a user based on the user's emotional state predicted from the article the user writes. We analyze the association between user-generated text and music by using a real-world dataset with user, text, music tripartite information collected from the social blogging website LiveJournal. The audio information represents various perceptual dimensions of music listening, including danceability, loudness, mode, and tempo; the emotional text information consists of bag-of-words and three dimensional affective states within an article: valence, arousal and dominance. To combine these factors for music recommendation, a factorization machine-based approach is taken. Our evaluation shows that the emotional context information mined from user-generated articles does improve the quality of recommendation, comparing to either the collaborative filtering approach or the content-based approach.

This paper proposes a music recommendation approach based on various similarity information via Factorization Machines (FM). We introduce the idea of similarity, which has been widely studied in the filed of information retrieval, and incorporate multiple feature similarities into the FM framework, including content-based and context-based similarities. The similarity information not only captures the similar patterns from the referred objects, but enhances the convergence speed and accuracy of FM. In addition, in order to avoid the noise within large similarity of features, we also adopt the grouping FM as an extended method to model the problem. In our experiments, a music-recommendation dataset is used to assess the performance of the proposed approach. The datasets is collected from an online blogging Web site, which includes user listening history, user profiles, social information, and music information. Our experimental results show that, with various types of feature similarities the performance of music recommendation can be enhanced significantly. Furthermore, via the grouping technique, the performance can be improved significantly in terms of Mean Average Precision, compared to the traditional collaborative filtering approach.

A scientific understanding of emotion experience requires information on the contexts in which the emotion is induced. Moreover, as one of the primary functions of music is to regulate the listener's mood, the individual's short-term music preference may reveal the emotional state of the individual. In light of these observations, this paper presents the first scientific study that exploits the online repository of social data to investigate the connections between a blogger's emotional state, user context manifested in the blog articles, and the content of the music titles the blogger attached to the post. A number of computational models are developed to evaluate the accuracy of different content or context cues in predicting emotional state, using 40,000 pieces of music listening records collected from the social blogging website LiveJournal. Our study shows that it is feasible to computationally model the latent structure underlying music listening and mood regulation. The average area under the receiver operating characteristic curve (AUC) for the content-based and context-based models attains 0.5462 and 0.6851, respectively. The association among user mood, music emotion, and individual's personality is also identified.

Thanks to the development of music audio analysis, state-of-the-art techniques can now detect musical attributes such as timbre, rhythm, and pitch with certain level of reliability and effectiveness. An emerging body of research has begun to model the high-level perceptual properties of music listening, including the mood and the preferable listening context of a music piece. Towards this goal, we propose a novel text-like feature representation that encodes the rich and time-varying information of music using a composite of features extracted from the song lyrics and audio signals. In particular, we investigate dictionary learning algorithms to optimize the generation of local feature descriptors and also probabilistic topic models to group semantically relevant text and audio words. This text-like representation leads to significant improvement in automatic mood classification over conventional audio features.