Diffusion-based audio and music generation models commonly generate music by constructing an image representation of audio (e.g., a mel-spectrogram) and then converting it to audio using a phase reconstruction model or vocoder. Typical vocoders, however, produce monophonic audio at lower resolutions (e.g., 16-24 kHz), which limits their effectiveness. We propose MusicHiFi -- an efficient high-fidelity stereophonic vocoder. Our method employs a cascade of three generative adversarial networks (GANs) that convert low-resolution mel-spectrograms to audio, upsamples to high-resolution audio via bandwidth expansion, and upmixes to stereophonic audio. Compared to previous work, we propose 1) a unified GAN-based generator and discriminator architecture and training procedure for each stage of our cascade, 2) a new fast, near downsampling-compatible bandwidth extension module, and 3) a new fast downmix-compatible mono-to-stereo upmixer that ensures the preservation of monophonic content in the output. We evaluate our approach using both objective and subjective listening tests and find our approach yields comparable or better audio quality, better spatialization control, and significantly faster inference speed compared to past work. Sound examples are at https://MusicHiFi.github.io/web/.
Multi-pitch estimation is a decades-long research problem involving the detection of pitch activity associated with concurrent musical events within multi-instrument mixtures. Supervised learning techniques have demonstrated solid performance on more narrow characterizations of the task, but suffer from limitations concerning the shortage of large-scale and diverse polyphonic music datasets with multi-pitch annotations. We present a suite of self-supervised learning objectives for multi-pitch estimation, which encourage the concentration of support around harmonics, invariance to timbral transformations, and equivariance to geometric transformations. These objectives are sufficient to train an entirely convolutional autoencoder to produce multi-pitch salience-grams directly, without any fine-tuning. Despite training exclusively on a collection of synthetic single-note audio samples, our fully self-supervised framework generalizes to polyphonic music mixtures, and achieves performance comparable to supervised models trained on conventional multi-pitch datasets.
Despite recent improvements in audio-text modeling, audio-text contrastive models still lag behind their image-text counterparts in scale and performance. We propose a method to improve both the scale and the training of audio-text contrastive models. Specifically, we craft a large-scale audio-text dataset consisting of over 13,000 hours of text-labeled audio, aided by large language model (LLM) processing and audio captioning. Further, we employ an masked autoencoder (MAE) pre-pretraining phase with random patch dropout, which allows us to both scale unlabeled audio datasets and train efficiently with variable length audio. After MAE pre-pretraining of our audio encoder, we train a contrastive model with an auxiliary captioning objective. Our final model, which we name Cacophony, achieves state-of-the-art performance on audio-text retrieval tasks, and exhibits competitive results on other downstream tasks such as zero-shot classification.
Dimensional representations of speech emotions such as the arousal-valence (AV) representation provide a continuous and fine-grained description and control than their categorical counterparts. They have wide applications in tasks such as dynamic emotion understanding and expressive text-to-speech synthesis. Existing methods that predict the dimensional emotion representation from speech cast it as a supervised regression task. These methods face data scarcity issues, as dimensional annotations are much harder to acquire than categorical labels. In this work, we propose to learn the AV representation from categorical emotion labels of speech. We start by learning a rich and emotion-relevant high-dimensional speech feature representation using self-supervised pre-training and emotion classification fine-tuning. This representation is then mapped to the 2D AV space according to psychological findings through anchored dimensionality reduction. Experiments show that our method achieves a Concordance Correlation Coefficient (CCC) performance comparable to state-of-the-art supervised regression methods on IEMOCAP without leveraging ground-truth AV annotations during training. This validates our proposed approach on AV prediction. Furthermore, visualization of AV predictions on MEAD and EmoDB datasets shows the interpretability of the learned AV representations.
Audio diffusion models can synthesize a wide variety of sounds. Existing models often operate on the latent domain with cascaded phase recovery modules to reconstruct waveform. This poses challenges when generating high-fidelity audio. In this paper, we propose EDMSound, a diffusion-based generative model in spectrogram domain under the framework of elucidated diffusion models (EDM). Combining with efficient deterministic sampler, we achieved similar Fr\'echet audio distance (FAD) score as top-ranked baseline with only 10 steps and reached state-of-the-art performance with 50 steps on the DCASE2023 foley sound generation benchmark. We also revealed a potential concern regarding diffusion based audio generation models that they tend to generate samples with high perceptual similarity to the data from training data. Project page: https://agentcooper2002.github.io/EDMSound/
Guitar tablature is a form of music notation widely used among guitarists. It captures not only the musical content of a piece, but also its implementation and ornamentation on the instrument. Guitar Tablature Transcription (GTT) is an important task with broad applications in music education and entertainment. Existing datasets are limited in size and scope, causing state-of-the-art GTT models trained on such datasets to suffer from overfitting and to fail in generalization across datasets. To address this issue, we developed a methodology for synthesizing SynthTab, a large-scale guitar tablature transcription dataset using multiple commercial acoustic and electric guitar plugins. This dataset is built on tablatures from DadaGP, which offers a vast collection and the degree of specificity we wish to transcribe. The proposed synthesis pipeline produces audio which faithfully adheres to the original fingerings, styles, and techniques specified in the tablature with diverse timbre. Experiments show that pre-training state-of-the-art GTT model on SynthTab improves transcription accuracy in same-dataset tests. More importantly, it significantly mitigates overfitting problems of GTT models in cross-dataset evaluation.
The rise of singing voice synthesis presents critical challenges to artists and industry stakeholders over unauthorized voice usage. Unlike synthesized speech, synthesized singing voices are typically released in songs containing strong background music that may hide synthesis artifacts. Additionally, singing voices present different acoustic and linguistic characteristics from speech utterances. These unique properties make singing voice deepfake detection a relevant but significantly different problem from synthetic speech detection. In this work, we propose the singing voice deepfake detection task. We first present SingFake, the first curated in-the-wild dataset consisting of 28.93 hours of bonafide and 29.40 hours of deepfake song clips in five languages from 40 singers. We provide a train/val/test split where the test sets include various scenarios. We then use SingFake to evaluate four state-of-the-art speech countermeasure systems trained on speech utterances. We find these systems lag significantly behind their performance on speech test data. When trained on SingFake, either using separated vocal tracks or song mixtures, these systems show substantial improvement. However, our evaluations also identify challenges associated with unseen singers, communication codecs, languages, and musical contexts, calling for dedicated research into singing voice deepfake detection. The SingFake dataset and related resources are available online.
Individualized head-related transfer functions (HRTFs) are crucial for accurate sound positioning in virtual auditory displays. As the acoustic measurement of HRTFs is resource-intensive, predicting individualized HRTFs using machine learning models is a promising approach at scale. Training such models require a unified HRTF representation across multiple databases to utilize their respectively limited samples. However, in addition to differences on the spatial sampling locations, recent studies have shown that, even for the common location, HRTFs across databases manifest consistent differences that make it trivial to tell which databases they come from. This poses a significant challenge for learning a unified HRTF representation across databases. In this work, we first identify the possible causes of these cross-database differences, attributing them to variations in the measurement setup. Then, we propose a novel approach to normalize the frequency responses of HRTFs across databases. We show that HRTFs from different databases cannot be classified by their database after normalization. We further show that these normalized HRTFs can be used to learn a more unified HRTF representation across databases than the prior art. We believe that this normalization approach paves the road to many data-intensive tasks on HRTF modeling.
In this paper, we aim to address the problem of channel robustness in speech countermeasure (CM) systems, which are used to distinguish synthetic speech from human natural speech. On the basis of two hypotheses, we suggest an approach for perturbing phase information during the training of time-domain CM systems. Communication networks often employ lossy compression codec that encodes only magnitude information, therefore heavily altering phase information. Also, state-of-the-art CM systems rely on phase information to identify spoofed speech. Thus, we believe the information loss in the phase domain induced by lossy compression codec degrades the performance of the unseen channel. We first establish the dependence of time-domain CM systems on phase information by perturbing phase in evaluation, showing strong degradation. Then, we demonstrated that perturbing phase during training leads to a significant performance improvement, whereas perturbing magnitude leads to further degradation.