Despite advances in diffusion-based text-to-music (TTM) methods, efficient, high-quality generation remains a challenge. We introduce Presto!, an approach to inference acceleration for score-based diffusion transformers via reducing both sampling steps and cost per step. To reduce steps, we develop a new score-based distribution matching distillation (DMD) method for the EDM-family of diffusion models, the first GAN-based distillation method for TTM. To reduce the cost per step, we develop a simple, but powerful improvement to a recent layer distillation method that improves learning via better preserving hidden state variance. Finally, we combine our step and layer distillation methods together for a dual-faceted approach. We evaluate our step and layer distillation methods independently and show each yield best-in-class performance. Our combined distillation method can generate high-quality outputs with improved diversity, accelerating our base model by 10-18x (230/435ms latency for 32 second mono/stereo 44.1kHz, 15x faster than the comparable SOTA model) — the fastest TTM to our knowledge.

We propose Diffusion Inference-Time T-Optimization (DITTO), a general-purpose framework for controlling pre-trained text-to-music diffusion models at inference-time via optimizing initial noise latents. Our method can be used to optimize through any differentiable feature matching loss to achieve a target (stylized) output and leverages gradient checkpointing for memory efficiency. We demonstrate a surprisingly wide-range of applications for music generation including inpainting, outpainting, and looping as well as intensity, melody, and musical structure control – all without ever fine-tuning the underlying model. When we compare our approach against related training, guidance, and optimization-based methods, we find DITTO achieves state-of-the-art performance on nearly all tasks, including outperforming comparable approaches on controllability, audio quality, and computational efficiency, thus opening the door for high-quality, flexible, training-free control of diffusion models. Sound examples can be found at https://ditto-music.github.io/web/.

Diffusion-based audio and music generation models commonly perform generation by constructing an image representation of audio (e.g., a mel-spectrogram) and then convert it to waveform using a phase reconstruction model or vocoder. Typical vocoders, however, produce monophonic audio at lower resolutions (e.g., 16-24 kHz), which limits their usefulness. 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 extension, and upmixes to stereophonic audio. Compared to past 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 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.

Text-to-music generation models are now capable of generating high-quality music audio in broad styles. However, text control is primarily suitable for the manipulation ofglobalmusical attributes like genre, mood, and tempo, and is less suitable for precise control overtime-varyingattributes such as the positions of beats in time or the changing dynamics of the music. We propose Music ControlNet, a diffusion-based music generation model that offers multiple precise, time-varying controls over generated audio. To imbue text-to-music models with time-varying control, we propose an approach analogous to pixel-wise control of the image-domain ControlNet method. Specifically, we extract controls from training audio yielding paired data, and fine-tune a diffusion-based conditional generative model over audio spectrograms given melody, dynamics, and rhythm controls. While the image-domain Uni-ControlNet method already allows generation with any subset of controls, we devise a new masking strategy to allow creators to input controls that are only partially specified in time. We evaluate both on controls extracted from audio and controls we expect creators to provide, demonstrating that we can generate realistic music that corresponds to control inputs in both settings. While few comparable music generation models exist, we benchmark against MusicGen, a recent model that accepts text and melody input, and show that our model generates music that is 49% more faithful to input melodies despite having 35x fewer parameters, training on 11x less data, and enabling two additional forms of time-varying control. Sound examples can be found athttps://musiccontrolnet.github.io/web/.

Adaptive filtering algorithms are pervasive throughout signal processing and have had a material impact on a wide variety of domains including audio processing, telecommunications, biomedical sensing, astrophysics and cosmology, seismology, and many more. Adaptive filters typically operate via specialized online, iterative optimization methods such as least-mean squares or recursive least squares and aim to process signals in unknown or nonstationary environments. Such algorithms, however, can be slow and laborious to develop, require domain expertise to create, and necessitate mathematical insight for improvement. In this work, we seek to improve upon hand-derived adaptive filter algorithms and present a comprehensive framework for learning online, adaptive signal processing algorithms or update rules directly from data. To do so, we frame the development of adaptive filters as a meta-learning problem in the context of deep learning and use a form of self-supervision to learn online iterative update rules for adaptive filters. To demonstrate our approach, we focus on audio applications and systematically develop meta-learned adaptive filters for five canonical audio problems including system identification, acoustic echo cancellation, blind equalization, multi-channel dereverberation, and beamforming.We compare our approach against common baselines and/or recent state-of-the-art methods. We show we can learn high-performing adaptive filters that operate in real-time and, in most cases, significantly outperform each method we compare against – all using a single general-purpose configuration of our approach.

The task of manipulating the level and/or effects of individual instruments to recompose a mixture of recordings, or remixing, is common across a variety of applications such as music production, audio-visual post-production, podcasts, and more. This process, however, traditionally requires access to individual source recordings, restricting the creative process. To work around this, source separation algorithms can separate a mixture into its respective components. Then, a user can adjust their levels and mix them back together. This two-step approach, however, still suffers from audible artifacts and motivates further work. In this work, we re-purpose Conv-TasNet, a well-known source separation model, into two neural remixing architectures that learn to remix directly rather than just to separate sources. We use an explicit loss term that directly measures remix quality and jointly optimize it with a separation loss. We evaluate our methods using the Slakh and MUSDB18 datasets and report remixing performance as well as the impact on source separation as a byproduct. Our results suggest that learning-to-remix significantly outperforms a strong separation baseline and is particularly useful for small volume changes.

Supervised learning for audio classification typically imposes a fixed class vocabulary, which can be limiting for real-world applications where the target class vocabulary is not known a priori or changes dynamically. In this work, we introduce a few-shot continual learning framework for audio classification, where we can continuously expand a trained base classifier to recognize novel classes based on only few labeled data at inference time. This enables fast and interactive model updates by end-users with minimal human effort. To do so, we leverage the dynamic few-shot learning technique and adapt it to a challenging multi-label audio classification scenario. We incorporate a recent state-of-the-art audio feature extraction model as a backbone and perform a comparative analysis of our approach on two popular audio datasets (ESC-50 and AudioSet). We conduct an in-depth evaluation to illustrate the complexities of the problem and show that, while there is still room for improvement, our method outperforms three baselines on novel class detection while maintaining its performance on base classes.

Text-based speech editors expedite the process of editing speech recordings by permitting editing via intuitive cut, copy, and paste operations on a speech transcript. A major drawback of current systems, however, is that edited recordings often sound unnatural because of prosody mismatches around edited regions. In our work, we propose a new context-aware method for more natural sounding text-based editing of speech. To do so, we 1) use a series of neural networks to generate salient prosody features that are dependent on the prosody of speech surrounding the edit and amenable to fine-grained user control 2) use the generated features to control a standard pitch-shift and time-stretch method and 3) apply a denoising neural network to remove artifacts induced by the signal manipulation to yield a high-fidelity result. We evaluate our approach using a subjective listening test, provide a detailed comparative analysis, and conclude several interesting insights.

We present a data-driven approach to automate audio signal processing by incorporating stateful third-party, audio effects as layers within a deep neural network. We then train a deep encoder to analyze input audio and control effect parameters to perform the desired signal manipulation, requiring only input-target paired audio data as supervision. To train our network with non-differentiable black-box effects layers, we use a fast, parallel stochastic gradient approximation scheme within a standard auto differentiation graph, yielding efficient end-to-end backpropagation. We demonstrate the power of our approach with three separate automatic audio production applications: tube amplifier emulation, automatic removal of breaths and pops from voice recordings, and automatic music mastering. We validate our results with a subjective listening test, showing our approach not only can enable new automatic audio effects tasks, but can yield results comparable to a specialized, state-of-the-art commercial solution for music mastering.

The reverberation time (T60) and the direct-to-reverberant ratio (DRR) are commonly used to characterize room acoustic environments. Both parameters can be measured from an acoustic impulse response (AIR) or using blind estimation methods that perform estimation directly from speech. When neural networks are used for blind estimation, however, a large realistic dataset is needed, which is expensive and time consuming to collect. To address this, we propose an AIR augmentation method that can parametrically control the T60 and DRR, allowing us to expand a small dataset of real AIRs into a balanced dataset orders of magnitude larger. Using this method, we train a previously proposed convolutional neural network (CNN) and show we can outperform past single-channel state-of-the-art methods. We then propose a more efficient, straightforward baseline CNN that is 4-5x faster, which provides an additional improvement and is better or comparable to all previously reported single- and multi-channel state-of-the-art methods.

Music similarity search is useful for a variety of creative tasks such as replacing one music recording with another recording with a similar "feel", a common task in video editing. For this task, it is typically necessary to define a similarity metric to compare one recording to another. Music similarity, however, is hard to define and depends on multiple simultaneous notions of similarity (i.e. genre, mood, instrument, tempo). While prior work ignore this issue, we embrace this idea and introduce the concept of multidimensional similarity and unify both global and specialized similarity metrics into a single, semantically disentangled multidimensional similarity metric. To do so, we adapt a variant of deep metric learning called conditional similarity networks to the audio domain and extend it using track-based information to control the specificity of our model. We evaluate our method and show that our single, multidimensional model outperforms both specialized similarity spaces and alternative baselines. We also run a user-study and show that our approach is favored by human annotators as well.

Audio production is a difficult process for many people], [and properly manipulating sound to achieve a certain effect is non-trivial. In this paper], [we present a method that facilitates this process by inferring appropriate audio effect parameters in order to make an input recording sound similar to an unrelated reference recording. We frame our work as a form of parametric style transfer that], [by design], [leverages existing audio production semantics and manipulation algorithms], [avoiding several issues that have plagued audio style transfer algorithms in the past. To demonstrate our approach], [we consider the task of controlling a parametric], [four-band infinite impulse response equalizer and show that we are able to predict the parameters necessary to transform the equalization style of one recording to another. The framework we present], [however], [is applicable to a wider range of parametric audio effects.

Locating perceptually similar sound events within a continuous recording is a common task for various audio applications. However, current tools require users to manually listen to and label all the locations of the sound events of interest, which is tedious and time-consuming. In this work, we (1) adapt state-of-the-art metric-based few-shot learning methods to automate the detection of similar-sounding events, requiring only one or few examples of the target event, (2) develop a method to automatically construct a partial set of labeled examples (negative samples) to reduce user labeling effort, and (3) develop an inference-time data augmentation method to increase detection accuracy. To validate our approach, we perform extensive comparative analysis of few-shot learning methods for the task of keyword detection in speech. We show that our approach successfully adapts closed-set few-shot learning approaches to an open-set sound event detection problem.

Many audio processing tasks require perceptual assessment.The "gold standard" of obtaining human judgments is timeconsuming, expensive, and cannot be used as an optimization criterion.On the other hand, automated metrics are efficient to compute but often correlate poorly with human judgment, particularly for audio differences at the threshold of human detection.In this work, we construct a metric by fitting a deep neural network to a new large dataset of crowdsourced human judgments.Subjects are prompted to answer a straightforward, objective question: are two recordings identical or not?These pairs are algorithmically generated under a variety of perturbations, including noise, reverb, and compression artifacts; the perturbation space is probed with the goal of efficiently identifying the just-noticeable difference (JND) level of the subject.We show that the resulting learned metric is well-calibrated with human judgments, outperforming baseline methods.Since it is a deep network, the metric is differentiable, making it suitable as a loss function for other tasks.Thus, simply replacing an existing loss (e.g., deep feature loss) with our metric yields significant improvement in a denoising network, as measured by subjective pairwise comparison.

Speech synthesis has recently seen significant improvements in fidelity, driven by the advent of neural vocoders and neural prosody generators. However, these systems lack intuitive user controls over prosody, making them unable to rectify prosody errors (e.g., misplaced emphases and contextually inappropriate emotions) or generate prosodies with diverse speaker excitement levels and emotions. We address these limitations with a user-controllable, context-aware neural prosody generator. Given a real or synthesized speech recording, our model allows a user to input prosody constraints for certain time frames and generates the remaining time frames from input text and contextual prosody. We also propose a pitch-shifting neural vocoder to modify input speech to match the synthesized prosody. Through objective and subjective evaluations we show that we can successfully incorporate user control into our prosody generation model without sacrificing the overall naturalness of the synthesized speech.

We present a new method to capture the acoustic characteristics of real-world rooms using commodity devices, and use the captured characteristics to generate similar sounding sources with virtual models. Given the captured audio and an approximate geometric model of a real-world room, we present a novel learning-based method to estimate its acoustic material properties. Our approach is based on deep neural networks that estimate the reverberation time and equalization of the room from recorded audio. These estimates are used to compute material properties related to room reverberation using a novel material optimization objective. We use the estimated acoustic material characteristics for audio rendering using interactive geometric sound propagation and highlight the performance on many real-world scenarios. We also perform a user study to evaluate the perceptual similarity between the recorded sounds and our rendered audio.

Traditional audio editing tools do not facilitate the task of separating a single mixture recording (e.g. pop song) into its respective sources (e.g. drums, vocal, etc.). Such ability, however, would be very useful for a wide variety of audio applications such as music remixing, audio denoising, and audio-based forensics. To address this issue, we present ISSE - an interactive source separation editor. ISSE is a new open-source, freely available, and cross-platform audio editing tool that enables a user to perform source separation by painting on time-frequency visualizations of sound, resulting in an interactive machine learning system. The system brings to life our previously proposed interaction paradigm and separation algorithm that learns from user-feedback to perform separation. For evaluation, we conducted user studies and compared results between inexperienced and expert users. For a variety of real-world tasks, we found that inexperienced users can achieve good separation quality with minimal instruction and expert users can achieve state-of-the-art separation quality.

We propose an interactive refinement method for supervised and semi-supervised single-channel source separation. The refinement method allows end-users to provide feedback to the separation process by painting on spectrogram displays of intermediate output results. The time-frequency annotations are then used to update the separation estimates and iteratively refine the results. The initial separation is performed using probabilistic latent component analysis and is then extended to incorporate the painting annotations using linear grouping expectation constraints via the framework of posterior regularization. Using a prototype user-interface, we show that the method is able to perform high-quality separation with minimal user-interaction.

In applications such as audio denoising, music transcription, music remixing, and audio-based forensics, it is desirable to decompose a single-channel recording into its respective sources. One of the current most effective class of methods to do so is based on non-negative matrix factorization and related latent variable models. Such techniques, however, typically perform poorly when no isolated training data is given and do not allow user feedback to correct for poor results. To overcome these issues, we allow a user to interactively constrain a latent variable model by painting on a time-frequency display of sound to guide the learning process. The annotations are used within the framework of posterior regularization to impose linear grouping constraints that would otherwise be difficult to achieve via standard priors. For the constraints considered, an efficient expectation-maximization algorithm is derived with closed-form multiplicative updates, drawing connections to non-negative matrix factorization methods, and allowing for high-quality interactive-rate separation without explicit training data.

We propose a method to both identify and synchronize multi-camera video recordings within a large collection of video and/or audio files. Landmark-based audio fingerprinting is used to match multiple recordings of the same event together and time-synchronize each file within the groups. Compared to prior work, we offer improvements towards event identification and a new synchronization refinement method that resolves inconsistent estimates and allows non-overlapping content to be synchronized within larger groups of recordings. Furthermore, the audio fingerprinting-based synchronization is shown to be equivalent to an efficient and scalable time-difference-of-arrival method using cross-correlation performed on a non-linearly transformed signal.

The behavior of the genetic algorithm (GA), a popular approach to search and optimization problems, is known to depend, among other factors, on the fitness function formula, the recombination operator, and the mutation operator. What has received less attention is the impact of the mating strategy that selects the chromosomes to be paired for recombination. Existing GA implementations mostly choose them probabilistically, according to their fitness function values, but we show that more sophisticated mating strategies can not only accelerate the search, but perhaps even improve the quality of the GA-generated solution. In our implementation, we took inspiration from the "opposites-attract” principle that is so common in nature. As a testbed, we chose the problem of 1-NN classifier tuning where genetic solutions have been employed before, and are thus well-understood by the research community. We propose three "instinct-based” mating strategies and experimentally investigate their behaviors.