We propose a new method for training a supervised source separation system that aims to learn the interdependent relationships between all combinations of sources in a mixture. Rather than independently estimating each source from a mix, we reframe the source separation problem as an Orderless Neural Autoregressive Density Estimator (NADE), and estimate each source from both the mix and a random subset of the other sources. We adapt a standard source separation architecture, Demucs, with additional inputs for each individual source, in addition to the input mixture. We randomly mask these input sources during training so that the network learns the conditional dependencies between the sources. By pairing this training method with a blocked Gibbs sampling procedure at inference time, we demonstrate that the network can iteratively improve its separation performance by conditioning a source estimate on its earlier source estimates. Experiments on two source separation datasets show that training a Demucs model with an Orderless NADE approach and using Gibbs sampling (up to 512 steps) at inference time strongly outperforms a Demucs baseline that uses a standard regression loss and direct (one step) estimation of sources.

Real-world data is high-dimensional: a book, image, or musical performance can easily contain hundreds of thousands of elements even after compression. However, the most commonly used autoregressive models, Transformers, are prohibitively expensive to scale to the number of inputs and layers needed to capture this long-range structure. We develop Perceiver AR, an autoregressive, modality-agnostic architecture which uses cross-attention to map long-range inputs to a small number of latents while also maintaining end-to-end causal masking. Perceiver AR can directly attend to over a hundred thousand tokens, enabling practical long-context density estimation without the need for hand-crafted sparsity patterns or memory mechanisms. When trained on images or music, Perceiver AR generates outputs with clear long-term coherence and structure. Our architecture also obtains state-of-the-art likelihood on long-sequence benchmarks, including 64x64 ImageNet images and PG-19 books.

There is an increasing interest from ML and HCI communities in empowering creators with better generative models and more intuitive interfaces with which to control them. In music, ML researchers have focused on training models capable of generating pieces with increasing long-range structure and musical coherence, while HCI researchers have separately focused on designing steering interfaces that support user control and ownership. In this study, we investigate how developments in both models and user interfaces are important for empowering co-creation where the goal is to create music that communicates particular imagery or ideas (e.g., as is common for other purposeful tasks in music creation like establishing mood or creating accompanying music for another media). Our study is distinguished in that it measures communication through both composer’s self-reported experiences, and how listeners evaluate this communication through the music. In an evaluation study with 26 composers creating 100+ pieces of music and listeners providing 1000+ head-to-head comparisons, we find that more expressive models and more steerable interfaces are important and complementary ways to make a difference in composers communicating through music and supporting their creative empowerment.

We consider the problem of learning high-level controls over the global structure of generated sequences, particularly in the context of symbolic music generation with complex language models. In this work, we present the Transformer autoencoder, which aggregates encodings of the input data across time to obtain a global representation of style from a given performance. We show it is possible to combine this global representation with other temporally distributed embeddings, enabling improved control over the separate aspects of performance style and melody. Empirically, we demonstrate the effectiveness of our method on various music generation tasks on the MAESTRO dataset and a YouTube dataset with 10,000+ hours of piano performances, where we achieve improvements in terms of log-likelihood and mean listening scores as compared to baselines.

We explore models for translating abstract musical ideas (scores, rhythms) into expressive performances using seq2seq and recurrent variational information bottleneck (VIB) models. Though seq2seq models usually require painstakingly aligned corpora, we show that it is possible to adapt an approach from the Generative Adversarial Network (GAN) literature (e.g. Pix2Pix, Vid2Vid) to sequences, creating large volumes of paired data by performing simple transformations and training generative models to plausibly invert these transformations. Music, and drumming in particular, provides a strong test case for this approach because many common transformations (quantization, removing voices) have clear semantics, and learning to invert them has real-world applications. Focusing on the case of drum set players, we create and release a new dataset for this purpose, containing over 13 hours of recordings by professional drummers aligned with fine-grained timing and dynamics information. We also explore some of the creative potential of these models, demonstrating improvements on state-of-the-art methods for Humanization (instantiating a performance from a musical score).

We provide an encoding and decoding strategy for efficient storage of analog data onto an array of Phase-Change Memory (PCM) devices. The PCM array is treated as an analog channel, with the stochastic relationship between write voltage and read resistance for each device determining its theoretical capacity. The encoder and decoder are implemented as neural networks with parameters that are trained end-to-end to minimize distortion for a fixed number of devices. To minimize distortion, the encoder and decoder must adapt jointly to the statistics of images and the statistics of the channel. Similar to Balle et al. (2017), we find that incorporating divisive normalization in the encoder, paired with de-normalization in the decoder, improves model performance. We show that the autoencoder achieves a rate-distortion performance above that achieved by a separate JPEG source coding and binary channel coding scheme. These results demonstrate the feasibility of exploiting the full analog dynamic range of PCM or other emerging memory devices for efficient storage of analog image data.

The Variational Autoencoder (VAE) has proven to be an effective model for producing semantically meaningful latent representations for natural data. However, it has thus far seen limited application to sequential data, and, as we demonstrate, existing recurrent VAE models have difficulty modeling sequences with long-term structure. To address this issue, we propose the use of a hierarchical decoder, which first outputs embeddings for subsequences of the input and then uses these embeddings to generate each subsequence independently. This structure encourages the model to utilize its latent code, thereby avoiding the "posterior collapse" problem which remains an issue for recurrent VAEs. We apply this architecture to modeling sequences of musical notes and find that it exhibits dramatically better sampling, interpolation, and reconstruction performance than a "flat" baseline model. An implementation of our "MusicVAE" is available online at https://goo.gl/magenta/musicvae-code.

Computer dialogue systems are designed with the intention of supporting meaningful interactions with humans. Common modes of communication include speech, text, and physical gestures. In this work we explore a communication paradigm in which the input and output channels consist of music. Specifically, we examine the musical interaction scenario of call and response. We present a system that utilizes a deep autoencoder to learn semantic embeddings of musical input. The system learns to transform these embeddings in a manner such that reconstructing from these transformation vectors produces appropriate musical responses. In order to generate a response the system employs a combination of generation and unit selection. Selection is based on a nearest neighbor search within the embedding space and for real-time application the search space is pruned using vector quantization. The live demo consists of a person playing a midi keyboard and the computer generating a response that is played through a loudspeaker.

Generative models in vision have seen rapid progress due to algorithmic improvements and the availability of high-quality image datasets. In this paper, we offer contributions in both these areas to enable similar progress in audio modeling. First, we detail a powerful new WaveNet-style autoencoder model that conditions an autoregressive decoder on temporal codes learned from the raw audio waveform. Second, we introduce NSynth, a large-scale and high-quality dataset of musical notes that is an order of magnitude larger than comparable public datasets. Using NSynth, we demonstrate improved qualitative and quantitative performance of the WaveNet autoencoder over a well-tuned spectral autoencoder baseline. Finally, we show that the model learns a manifold of embeddings that allows for morphing between instruments, meaningfully interpolating in timbre to create new types of sounds that are realistic and expressive.

We show that an end-to-end deep learning approach can be used to recognize either English or Mandarin Chinese speech–two vastly different languages. Because it replaces entire pipelines of hand-engineered components with neural networks, end-to-end learning allows us to handle a diverse variety of speech including noisy environments, accents and different languages. Key to our approach is our application of HPC techniques, enabling experiments that previously took weeks to now run in days. This allows us to iterate more quickly to identify superior architectures and algorithms. As a result, in several cases, our system is competitive with the transcription of human workers when benchmarked on standard datasets. Finally, using a technique called Batch Dispatch with GPUs in the data center, we show that our system can be inexpensively deployed in an online setting, delivering low latency when serving users at scale.

This paper introduces a new technique for mapping Deep Recurrent Neural Networks (RNN) efficiently onto GPUs. We show how it is possi- ble to achieve substantially higher computational throughput at low mini-batch sizes than direct implementations of RNNs based on matrix multiplications. The key to our approach is the use of persistent computational kernels that exploit the GPU’s inverted memory hierarchy to reuse network weights over multiple timesteps. Our initial implementation sustains 2.8 TFLOP/s at a mini-batch size of 4 on an NVIDIA TitanX GPU. This provides a 16x reduction in activation memory footprint, enables model training with 12x more parameters on the same hardware, allows us to strongly scale RNN training to 128 GPUs, and allows us to efficiently explore end-to-end speech recognition models with over 100 layers.

Deep learning has dramatically improved the performance of speech recognition systems through learning hierarchies of features optimized for the task at hand. However, true end-to-end learning, where features are learned directly from waveforms, has only recently reached the performance of hand-tailored representations based on the Fourier transform. In this paper, we detail an approach to use convolutional filters to push past the inherent tradeoff of temporal and frequency resolution that exists for spectral representations. At increased computational cost, we show that increasing temporal resolution via reduced stride and increasing frequency resolution via additional filters delivers significant performance improvements. Further, we find more efficient representations by simultaneously learning at multiple scales, leading to an overall decrease in word error rate on a difficult internal speech test set by 20.7% relative to networks with the same number of parameters trained on spectrograms.