Creating a Good Teacher for Knowledge Distillation in Acoustic Scene Classification

Knowledge Distillation (KD) is a widespread technique for compressing the knowledge of large models into more compact and efficient models. KD has proved to be highly effective in building well-performing low-complexity Acoustic Scene Classification (ASC) systems and was used in all the top-ranked submissions to this task of the annual DCASE challenge in the past three years. There is extensive research available on establishing the KD process, designing efficient student models, and forming well-performing teacher ensembles. However, less research has been conducted on investigating which teacher model attributes are beneficial for low-complexity students. In this work, we try to close this gap by studying the effects on the student's performance when using different teacher network architectures, varying the teacher model size, training them with different device generalization methods, and applying different ensembling strategies. The results show that teacher model sizes, device generalization methods, the ensembling strategy and the ensemble size are key factors for a well-performing student network.

Exploring Performance-Complexity Trade-Offs in Sound Event Detection

We target the problem of developing new low-complexity networks for the sound event detection task. Our goal is to meticulously analyze the performance-complexity trade-off, aiming to be competitive with the large state-of-the-art models, at a fraction of the computational requirements. We find that low-complexity convolutional models previously proposed for audio tagging can be effectively adapted for event detection (which requires frame-wise prediction) by adjusting convolutional strides, removing the global pooling, and, importantly, adding a sequence model before the (now frame-wise) classification heads. Systematic experiments reveal that the best choice for the sequence model type depends on which complexity metric is most important for the given application. We also investigate the impact of enhanced training strategies such as knowledge distillation. In the end, we show that combined with an optimized training strategy, we can reach event detection performance comparable to state-of-the-art transformers while requiring only around 5% of the parameters. We release all our pre-trained models and the code for reproducing this work to support future research in low-complexity sound event detection at https://github.com/theMoro/EfficientSED.

Estimating Musical Surprisal in Audio

In modeling musical surprisal expectancy with computational methods, it has been proposed to use the information content (IC) of one-step predictions from an autoregressive model as a proxy for surprisal in symbolic music. With an appropriately chosen model, the IC of musical events has been shown to correlate with human perception of surprise and complexity aspects, including tonal and rhythmic complexity. This work investigates whether an analogous methodology can be applied to music audio. We train an autoregressive Transformer model to predict compressed latent audio representations of a pretrained autoencoder network. We verify learning effects by estimating the decrease in IC with repetitions. We investigate the mean IC of musical segment types (e.g., A or B) and find that segment types appearing later in a piece have a higher IC than earlier ones on average. We investigate the IC's relation to audio and musical features and find it correlated with timbral variations and loudness and, to a lesser extent, dissonance, rhythmic complexity, and onset density related to audio and musical features. Finally, we investigate if the IC can predict EEG responses to songs and thus model humans' surprisal in music. We provide code for our method on github.com/sonycslparis/audioic.

Language Models for Music Medicine Generation

Music therapy has been shown in recent years to provide multiple health benefits related to emotional wellness. In turn, maintaining a healthy emotional state has proven to be effective for patients undergoing treatment, such as Parkinson's patients or patients suffering from stress and anxiety. We propose fine-tuning MusicGen, a music-generating transformer model, to create short musical clips that assist patients in transitioning from negative to desired emotional states. Using low-rank decomposition fine-tuning on the MTG-Jamendo Dataset with emotion tags, we generate 30-second clips that adhere to the iso principle, guiding patients through intermediate states in the valence-arousal circumplex. The generated music is evaluated using a music emotion recognition model to ensure alignment with intended emotions. By concatenating these clips, we produce a 15-minute "music medicine" resembling a music therapy session. Our approach is the first model to leverage Language Models to generate music medicine. Ultimately, the output is intended to be used as a temporary relief between music therapy sessions with a board-certified therapist.

Effective Pre-Training of Audio Transformers for Sound Event Detection

We propose a pre-training pipeline for audio spectrogram transformers for frame-level sound event detection tasks. On top of common pre-training steps, we add a meticulously designed training routine on AudioSet frame-level annotations. This includes a balanced sampler, aggressive data augmentation, and ensemble knowledge distillation. For five transformers, we obtain a substantial performance improvement over previously available checkpoints both on AudioSet frame-level predictions and on frame-level sound event detection downstream tasks, confirming our pipeline's effectiveness. We publish the resulting checkpoints that researchers can directly fine-tune to build high-performance models for sound event detection tasks.

Estimated Audio-Caption Correspondences Improve Language-Based Audio Retrieval

Dual-encoder-based audio retrieval systems are commonly optimized with contrastive learning on a set of matching and mismatching audio-caption pairs. This leads to a shared embedding space in which corresponding items from the two modalities end up close together. Since audio-caption datasets typically only contain matching pairs of recordings and descriptions, it has become common practice to create mismatching pairs by pairing the audio with a caption randomly drawn from the dataset. This is not ideal because the randomly sampled caption could, just by chance, partly or entirely describe the audio recording. However, correspondence information for all possible pairs is costly to annotate and thus typically unavailable; we, therefore, suggest substituting it with estimated correspondences. To this end, we propose a two-staged training procedure in which multiple retrieval models are first trained as usual, i.e., without estimated correspondences. In the second stage, the audio-caption correspondences predicted by these models then serve as prediction targets. We evaluate our method on the ClothoV2 and the AudioCaps benchmark and show that it improves retrieval performance, even in a restricting self-distillation setting where a single model generates and then learns from the estimated correspondences. We further show that our method outperforms the current state of the art by 1.6 pp. mAP@10 on the ClothoV2 benchmark.

Improving Query-by-Vocal Imitation with Contrastive Learning and Audio Pretraining

Query-by-Vocal Imitation (QBV) is about searching audio files within databases using vocal imitations created by the user's voice. Since most humans can effectively communicate sound concepts through voice, QBV offers the more intuitive and convenient approach compared to text-based search. To fully leverage QBV, developing robust audio feature representations for both the vocal imitation and the original sound is crucial. In this paper, we present a new system for QBV that utilizes the feature extraction capabilities of Convolutional Neural Networks pre-trained with large-scale general-purpose audio datasets. We integrate these pre-trained models into a dual encoder architecture and fine-tune them end-to-end using contrastive learning. A distinctive aspect of our proposed method is the fine-tuning strategy of pre-trained models using an adapted NT-Xent loss for contrastive learning, creating a shared embedding space for reference recordings and vocal imitations. The proposed system significantly enhances audio retrieval performance, establishing a new state of the art on both coarse- and fine-grained QBV tasks.

Controlling Surprisal in Music Generation via Information Content Curve Matching

In recent years, the quality and public interest in music generation systems have grown, encouraging research into various ways to control these systems. We propose a novel method for controlling surprisal in music generation using sequence models. To achieve this goal, we define a metric called Instantaneous Information Content (IIC). The IIC serves as a proxy function for the perceived musical surprisal (as estimated from a probabilistic model) and can be calculated at any point within a music piece. This enables the comparison of surprisal across different musical content even if the musical events occur in irregular time intervals. We use beam search to generate musical material whose IIC curve closely approximates a given target IIC. We experimentally show that the IIC correlates with harmonic and rhythmic complexity and note density. The correlation decreases with the length of the musical context used for estimating the IIC. Finally, we conduct a qualitative user study to test if human listeners can identify the IIC curves that have been used as targets when generating the respective musical material. We provide code for creating IIC interpolations and IIC visualizations on https://github.com/muthissar/iic.

Quantifying the Corpus Bias Problem in Automatic Music Transcription Systems

Automatic Music Transcription (AMT) is the task of recognizing notes in audio recordings of music. The State-of-the-Art (SotA) benchmarks have been dominated by deep learning systems. Due to the scarcity of high quality data, they are usually trained and evaluated exclusively or predominantly on classical piano music. Unfortunately, that hinders our ability to understand how they generalize to other music. Previous works have revealed several aspects of memorization and overfitting in these systems. We identify two primary sources of distribution shift: the music, and the sound. Complementing recent results on the sound axis (i.e. acoustics, timbre), we investigate the musical one (i.e. note combinations, dynamics, genre). We evaluate the performance of several SotA AMT systems on two new experimental test sets which we carefully construct to emulate different levels of musical distribution shift. Our results reveal a stark performance gap, shedding further light on the Corpus Bias problem, and the extent to which it continues to trouble these systems.

TheGlueNote: Learned Representations for Robust and Flexible Note Alignment

Note alignment refers to the task of matching individual notes of two versions of the same symbolically encoded piece. Methods addressing this task commonly rely on sequence alignment algorithms such as Hidden Markov Models or Dynamic Time Warping (DTW) applied directly to note or onset sequences. While successful in many cases, such methods struggle with large mismatches between the versions. In this work, we learn note-wise representations from data augmented with various complex mismatch cases, e.g. repeats, skips, block insertions, and long trills. At the heart of our approach lies a transformer encoder network - TheGlueNote - which predicts pairwise note similarities for two 512 note subsequences. We postprocess the predicted similarities using flavors of weightedDTW and pitch-separated onsetDTW to retrieve note matches for two sequences of arbitrary length. Our approach performs on par with the state of the art in terms of note alignment accuracy, is considerably more robust to version mismatches, and works directly on any pair of MIDI files.