Nicholas J. Bryan

Yatong Bai, Jonah Casebeer, Somayeh Sojoudi et al.

We present Distributional RewArds for Generative OptimizatioN (DRAGON), a versatile framework for fine-tuning media generation models towards a desired outcome. Compared with traditional reinforcement learning with human feedback (RLHF) or pairwise preference approaches such as direct preference optimization (DPO), DRAGON is more flexible. It can optimize reward functions that evaluate either individual examples or distributions of them, making it compatible with a broad spectrum of instance-wise, instance-to-distribution, and distribution-to-distribution rewards. Leveraging this versatility, we construct novel reward functions by selecting an encoder and a set of reference examples to create an exemplar distribution. When cross-modal encoders such as CLAP are used, the reference may be of a different modality (text versus audio). Then, DRAGON gathers online and on-policy generations, scores them with the reward function to construct a positive demonstration set and a negative set, and leverages the contrast between the two finite sets to approximate distributional reward optimization. For evaluation, we fine-tune an audio-domain text-to-music diffusion model with 20 reward functions, including a custom music aesthetics model, CLAP score, Vendi diversity, and Frechet audio distance (FAD). We further compare instance-wise (per-song) and full-dataset FAD settings while ablating multiple FAD encoders and reference sets. Over all 20 target rewards, DRAGON achieves an 81.45% average win rate. Moreover, reward functions based on exemplar sets enhance generations and are comparable to model-based rewards. With an appropriate exemplar set, DRAGON achieves a 60.95% human-voted music quality win rate without training on human preference annotations. DRAGON is a new approach to designing and optimizing reward functions for improving human-perceived quality. Demos at https://ml-dragon.github.io/web

Shih-Lun Wu, Ge Zhu, Juan-Pablo Caceres et al.

Music stem generation, the task of producing musically-synchronized and isolated instrument audio clips, offers the potential of greater user control and better alignment with musician workflows compared to conventional text-to-music models. Existing stem generation approaches, however, either rely on fixed architectures that output a predefined set of stems in parallel, or generate only one stem at a time, resulting in slow inference despite flexibility in stem combination. We propose Stemphonic, a diffusion-/flow-based framework that overcomes this trade-off and generates a variable set of synchronized stems in one inference pass. During training, we treat each stem as a batch element, group synchronized stems in a batch, and apply a shared noise latent to each group. At inference-time, we use a shared initial noise latent and stem-specific text inputs to generate synchronized multi-stem outputs in one pass. We further expand our approach to enable one-pass conditional multi-stem generation and stem-wise activity controls to empower users to iteratively generate and orchestrate the temporal layering of a mix. We benchmark our results on multiple open-source stem evaluation sets and show that Stemphonic produces higher-quality outputs while accelerating the full mix generation process by 25 to 50%. Demos at: https://stemphonic-demo.vercel.app.

Yan-Bo Lin, Jonah Casebeer, Long Mai et al.

Generating music that temporally aligns with video events is challenging for existing text-to-music models, which lack fine-grained temporal control. We introduce V2M-Zero, a zero-pair video-to-music generation approach that outputs time-aligned music for video. Our method is motivated by a key observation: temporal synchronization requires matching when and how much change occurs, not what changes. While musical and visual events differ semantically, they exhibit shared temporal structure that can be captured independently within each modality. We capture this structure through event curves computed from intra-modal similarity using pretrained music and video encoders. By measuring temporal change within each modality independently, these curves provide comparable representations across modalities. This enables a simple training strategy: fine-tune a text-to-music model on music-event curves, then substitute video-event curves at inference without cross-modal training or paired data. Across OES-Pub, MovieGenBench-Music, and AIST++, V2M-Zero achieves substantial gains over paired-data baselines: 5-21% higher audio quality, 13-15% better semantic alignment, 21-52% improved temporal synchronization, and 28% higher beat alignment on dance videos. We find similar results via a large crowd-source subjective listening test. Overall, our results validate that temporal alignment through within-modality features, rather than paired cross-modal supervision, is effective for video-to-music generation. Results are available at https://genjib.github.io/v2m_zero/

Irmak Bukey, Zhepei Wang, Chris Donahue et al.

Music captioning, or the task of generating a natural language description of music, is useful for both music understanding and controllable music generation. Training captioning models, however, typically requires high-quality music caption data which is scarce compared to metadata (e.g., genre, mood, etc.). As a result, it is common to use large language models (LLMs) to synthesize captions from metadata to generate training data for captioning models, though this process imposes a fixed stylization and entangles factual information with natural language style. As a more direct approach, we propose metadata-based captioning. We train a metadata prediction model to infer detailed music metadata from audio and then convert it into expressive captions via pre-trained LLMs at inference time. Compared to a strong end-to-end baseline trained on LLM-generated captions derived from metadata, our method: (1) achieves comparable performance in less training time over end-to-end captioners, (2) offers flexibility to easily change stylization post-training, enabling output captions to be tailored to specific stylistic and quality requirements, and (3) can be prompted with audio and partial metadata to enable powerful metadata imputation or in-filling--a common task for organizing music data.

Yu Wang, Nicholas J. Bryan, Justin Salamon et al.

Few-shot learning aims to train models that can recognize novel classes given just a handful of labeled examples, known as the support set. While the field has seen notable advances in recent years, they have often focused on multi-class image classification. Audio, in contrast, is often multi-label due to overlapping sounds, resulting in unique properties such as polyphony and signal-to-noise ratios (SNR). This leads to unanswered questions concerning the impact such audio properties may have on few-shot learning system design, performance, and human-computer interaction, as it is typically up to the user to collect and provide inference-time support set examples. We address these questions through a series of experiments designed to elucidate the answers to these questions. We introduce two novel datasets, FSD-MIX-CLIPS and FSD-MIX-SED, whose programmatic generation allows us to explore these questions systematically. Our experiments lead to audio-specific insights on few-shot learning, some of which are at odds with recent findings in the image domain: there is no best one-size-fits-all model, method, and support set selection criterion. Rather, it depends on the expected application scenario. Our code and data are available at https://github.com/wangyu/rethink-audio-fsl.

Zachary Novack, Julian McAuley, Taylor Berg-Kirkpatrick et al.

We propose Diffusion Inference-Time T-Optimization (DITTO), a general-purpose frame-work 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/.

Minz Won, Justin Salamon, Nicholas J. Bryan et al.

Content creators often use music to enhance their stories, as it can be a powerful tool to convey emotion. In this paper, our goal is to help creators find music to match the emotion of their story. We focus on text-based stories that can be auralized (e.g., books), use multiple sentences as input queries, and automatically retrieve matching music. We formalize this task as a cross-modal text-to-music retrieval problem. Both the music and text domains have existing datasets with emotion labels, but mismatched emotion vocabularies prevent us from using mood or emotion annotations directly for matching. To address this challenge, we propose and investigate several emotion embedding spaces, both manually defined (e.g., valence/arousal) and data-driven (e.g., Word2Vec and metric learning) to bridge this gap. Our experiments show that by leveraging these embedding spaces, we are able to successfully bridge the gap between modalities to facilitate cross modal retrieval. We show that our method can leverage the well established valence-arousal space, but that it can also achieve our goal via data-driven embedding spaces. By leveraging data-driven embeddings, our approach has the potential of being generalized to other retrieval tasks that require broader or completely different vocabularies.

Jonah Casebeer, Nicholas J. Bryan, Paris Smaragdis

Adaptive filtering algorithms are commonplace in signal processing and have wide-ranging applications from single-channel denoising to multi-channel acoustic echo cancellation and adaptive beamforming. Such algorithms typically operate via specialized online, iterative optimization methods and have achieved tremendous success, but require expert knowledge, are slow to develop, and are difficult to customize. In our work, we present a new method to automatically learn adaptive filtering update rules directly from data. To do so, we frame adaptive filtering as a differentiable operator and train a learned optimizer to output a gradient descent-based update rule from data via backpropagation through time. We demonstrate our general approach on an acoustic echo cancellation task (single-talk with noise) and show that we can learn high-performing adaptive filters for a variety of common linear and non-linear multidelayed block frequency domain filter architectures. We also find that our learned update rules exhibit fast convergence, can optimize in the presence of nonlinearities, and are robust to acoustic scene changes despite never encountering any during training.

Max Morrison, Zeyu Jin, Nicholas J. Bryan et al.

Modifying the pitch and timing of an audio signal are fundamental audio editing operations with applications in speech manipulation, audio-visual synchronization, and singing voice editing and synthesis. Thus far, methods for pitch-shifting and time-stretching that use digital signal processing (DSP) have been favored over deep learning approaches due to their speed and relatively higher quality. However, even existing DSP-based methods for pitch-shifting and time-stretching induce artifacts that degrade audio quality. In this paper, we propose Controllable LPCNet (CLPCNet), an improved LPCNet vocoder capable of pitch-shifting and time-stretching of speech. For objective evaluation, we show that CLPCNet performs pitch-shifting of speech on unseen datasets with high accuracy relative to prior neural methods. For subjective evaluation, we demonstrate that the quality and naturalness of pitch-shifting and time-stretching with CLPCNet on unseen datasets meets or exceeds competitive neural- or DSP-based approaches.

Haici Yang, Shivani Firodiya, Nicholas J. Bryan et al.

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 learn to remix music directly by re-purposing Conv-TasNet, a well-known source separation model, into two neural remixing architectures. To do this, 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.

Marco A. Martínez Ramírez, Oliver Wang, Paris Smaragdis et al.

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.

Max Morrison, Lucas Rencker, Zeyu Jin et al.

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.

Jongpil Lee, Nicholas J. Bryan, Justin Salamon et al.

Deep representation learning offers a powerful paradigm for mapping input data onto an organized embedding space and is useful for many music information retrieval tasks. Two central methods for representation learning include deep metric learning and classification, both having the same goal of learning a representation that can generalize well across tasks. Along with generalization, the emerging concept of disentangled representations is also of great interest, where multiple semantic concepts (e.g., genre, mood, instrumentation) are learned jointly but remain separable in the learned representation space. In this paper we present a single representation learning framework that elucidates the relationship between metric learning, classification, and disentanglement in a holistic manner. For this, we (1) outline past work on the relationship between metric learning and classification, (2) extend this relationship to multi-label data by exploring three different learning approaches and their disentangled versions, and (3) evaluate all models on four tasks (training time, similarity retrieval, auto-tagging, and triplet prediction). We find that classification-based models are generally advantageous for training time, similarity retrieval, and auto-tagging, while deep metric learning exhibits better performance for triplet-prediction. Finally, we show that our proposed approach yields state-of-the-art results for music auto-tagging.

Jongpil Lee, Nicholas J. Bryan, Justin Salamon et al.

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.

Max Morrison, Zeyu Jin, Justin Salamon et al.

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.

Yu Wang, Justin Salamon, Mark Cartwright et al.

Data-driven approaches to automatic drum transcription (ADT) are often limited to a predefined, small vocabulary of percussion instrument classes. Such models cannot recognize out-of-vocabulary classes nor are they able to adapt to finer-grained vocabularies. In this work, we address open vocabulary ADT by introducing few-shot learning to the task. We train a Prototypical Network on a synthetic dataset and evaluate the model on multiple real-world ADT datasets with polyphonic accompaniment. We show that, given just a handful of selected examples at inference time, we can match and in some cases outperform a state-of-the-art supervised ADT approach under a fixed vocabulary setting. At the same time, we show that our model can successfully generalize to finer-grained or extended vocabularies unseen during training, a scenario where supervised approaches cannot operate at all. We provide a detailed analysis of our experimental results, including a breakdown of performance by sound class and by polyphony.

Pranay Manocha, Adam Finkelstein, Richard Zhang et al.

Many audio processing tasks require perceptual assessment. The ``gold standard`` of obtaining human judgments is time-consuming, 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.

Zhenyu Tang, Nicholas J. Bryan, Dingzeyu Li et al.

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.

Nicholas J. Bryan

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.