Automatic music generation dates back to more than half a century.^{1, 2, 3, 4} A prominent approach is to generate music symbolically in the form of a piano roll, which specifies the timing, pitch, velocity, and instrument of each note to be played. This has led to impressive results like producing Bach chorals,^{5, 6} polyphonic music with multiple instruments,^{7, 8, 9} as well as minute long musical pieces.^{10, 11, 12}

But symbolic generators have limitations—they cannot capture human voices or many of the more subtle timbres, dynamics, and expressivity that are essential to music. A different approach^A is to model music directly as raw audio.^{13, 14, 15, 16} Generating music at the audio level is challenging since the sequences are very long.¹⁷ A typical 4-minute song at CD quality (44 kHz, 16-bit) has over 10 million timesteps. For comparison, GPT‑2 had 1,000 timesteps and OpenAI Five⁠ took tens of thousands of timesteps per game. Thus, to learn the high level semantics of music, a model would have to deal with extremely long-range dependencies.

One way of addressing the long input problem is to use an autoencoder that compresses raw audio to a lower-dimensional space by discarding some of the perceptually irrelevant bits of information. We can then train a model to generate audio in this compressed space, and upsample back to the raw audio space.^{25, 17}

We chose to work on music because we want to continue to push the boundaries of generative models. Our previous work on MuseNet⁠ explored synthesizing music based on large amounts of MIDI data. Now in raw audio, our models must learn to tackle high diversity as well as very long range structure, and the raw audio domain is particularly unforgiving of errors in short, medium, or long term timing.

Next, we train the prior models whose goal is to learn the distribution of music codes encoded by VQ-VAE and to generate music in this compressed discrete space. Like the VQ-VAE, we have three levels of priors: a top-level prior that generates the most compressed codes, and two upsampling priors that generate less compressed codes conditioned on above.

The top-level prior models the long-range structure of music, and samples decoded from this level have lower audio quality but capture high-level semantics like singing and melodies. The middle and bottom upsampling priors add local musical structures like timbre, significantly improving the audio quality.

We train these as autoregressive models using a simplified variant of Sparse Transformers.^{29, 30} Each of these models has 72 layers of factorized self-attention on a context of 8192 codes, which corresponds to approximately 24 seconds, 6 seconds, and 1.5 seconds of raw audio at the top, middle and bottom levels, respectively.

Once all of the priors are trained, we can generate codes from the top level, upsample them using the upsamplers, and decode them back to the raw audio space using the VQ-VAE decoder to sample novel songs.

To train this model, we crawled the web to curate a new dataset of 1.2 million songs (600,000 of which are in English), paired with the corresponding lyrics and metadata from LyricWiki⁠(opens in a new window). The metadata includes artist, album genre, and year of the songs, along with common moods or playlist keywords associated with each song. We train on 32-bit, 44.1 kHz raw audio, and perform data augmentation by randomly downmixing the right and left channels to produce mono audio.