# Jukebox (AI model)
Jukebox is a neural network that OpenAI announced in April 2020 which generates music, including rudimentary singing, as raw audio in a variety of genres and artist styles.[1](https://openai.com/index/jukebox/) It generated complete songs with vocals directly in the audio waveform, conditioned on artist, genre, and unaligned lyrics.[2](https://arxiv.org/pdf/2005.0341) OpenAI released the model weights and code together with a tool to explore generated samples.[1](https://openai.com/index/jukebox/)
| Key fact | Value |
|---|---|
| Announced | April 2020 (GitHub repository created 2020-04-29)[1](https://openai.com/index/jukebox/) • [3](https://github.com/openai/jukebox/) |
| Output | Raw 44.1 kHz audio with vocals, conditioned on artist, genre, and lyrics[1](https://openai.com/index/jukebox/) |
| Largest prior | 5 billion parameters, trained on 512 V100 GPUs for 4 weeks[2](https://arxiv.org/pdf/2005.0341) |
| Training data | 1.2 million songs (600,000 in English) with lyrics and metadata from LyricWiki[1](https://openai.com/index/jukebox/) |
| Generation speed | About 9 hours to fully render one minute of audio[1](https://openai.com/index/jukebox/) |
| License | Noncommercial Use License covering both code and weights; repository archived[3](https://github.com/openai/jukebox/) |
| Public samples | 7,116 non-cherry-picked generated songs in the official Sample Explorer[4](https://jukebox.openai.com/?song=802878241) |
## Architecture and training as published
Generating a song directly as a waveform means producing an enormous sequence. OpenAI's announcement noted that a typical 4-minute song at CD quality (44 kHz, 16-bit) has over 10 million timesteps, compared with the 1,000 timesteps of GPT-2.[1](https://openai.com/index/jukebox/) Jukebox handled this with a multi-scale VQ-VAE ([vector-quantized variational autoencoder](https://www.edgepedia.org/vector-quantized-variational-autoencoder)) that compresses raw audio into discrete codes, which autoregressive [Transformers](https://www.edgepedia.org/transformers) then model.[2](https://arxiv.org/pdf/2005.0341)
The VQ-VAE used three levels of compression, compressing the 44 kHz raw audio by 8x, 32x, and 128x, each with a codebook size of 2048.[1](https://openai.com/index/jukebox/) A context of 8192 codes corresponds to roughly 24 seconds of raw audio at the top level, 6 seconds at the middle level, and 1.5 seconds at the bottom level.[2](https://arxiv.org/pdf/2005.0341) The priors were autoregressive Sparse Transformers with 72 layers of factorized self-attention on a context of 8192 codes.[1](https://openai.com/index/jukebox/)
The compute disclosed in the paper was substantial and split across components. The VQ-VAE itself has 2 million parameters and was trained on 9-second audio clips on 256 V100 GPUs for 3 days. The upsamplers have one billion parameters, trained on 128 V100s for 2 weeks, and the top-level prior has 5 billion parameters, trained on 512 V100s for 4 weeks.[2](https://arxiv.org/pdf/2005.0341) Training data was a newly crawled dataset of 1.2 million songs, 600,000 of them in English, paired with lyrics and metadata from LyricWiki, all as 32-bit 44.1 kHz raw audio.[1](https://openai.com/index/jukebox/) The model was trained only on songs whose primary detected language is English.[2](https://arxiv.org/pdf/2005.0341)
Two engineering details addressed known failure modes. Random restarts reset codebook vectors when usage fell below a threshold, mitigating codebook collapse, and a spectral loss was added to help reconstruct higher frequencies.[1](https://openai.com/index/jukebox/) Lyric tokens and VQ-VAE codes were merged into a single larger vocabulary and flattened into one sequence, using attn_order=12 with prime_attention layers.[5](https://github.com/openai/jukebox/blob/master/README.md) The 5b and 5b_lyrics top-level priors occupy 10.3 GB and 11.5 GB of storage respectively, with peak transformer key/value cache memory of about 1 GB for the largest model.[5](https://github.com/openai/jukebox/blob/master/README.md)
## What the results actually showed
OpenAI's own assessment was that the combined model at scale can generate high-fidelity and diverse songs with coherence up to multiple minutes, steerable by artist and genre for musical and vocal style and by unaligned lyrics for more controllable singing.[2](https://arxiv.org/pdf/2005.0341) But the same sources state the limits plainly: there is a significant gap between these generations and human-created music; the samples lack traditional larger music structures such as choruses that repeat or melodies with a question-and-answer form; and audible noise or scratchiness appears at the smallest scale, introduced by the downsampling and upsampling pipeline.[2](https://arxiv.org/pdf/2005.0341) • [1](https://openai.com/index/jukebox/)
The public evidence base was unusually large for its time. OpenAI released thousands of non-cherry-picked samples along with weights and code,[2](https://arxiv.org/pdf/2005.0341) and the official Sample Explorer lists 7,116 generated songs, browsable by prior size, lyrics, genre, and artist conditioning.[4](https://jukebox.openai.com/?song=802878241) Reception among working musicians was cool: OpenAI shared Jukebox with an initial set of 10 musicians from various genres, who did not find it immediately applicable to their creative process given its current limitations.[1](https://openai.com/index/jukebox/)
## Availability, licensing and cost
Jukebox was released as code and weights on GitHub, where the repository was created on 2020-04-29 and is now archived with the status note that code is provided as-is with no updates expected.[3](https://github.com/openai/jukebox/) Both the released code and the weights are covered by a Noncommercial Use License, ruling out commercial use.[3](https://github.com/openai/jukebox/)
Speed was the practical barrier. OpenAI stated that it takes approximately 9 hours to fully render one minute of audio through the models, so they cannot yet be used in interactive applications.[1](https://openai.com/index/jukebox/) The paper breaks this down as around 1 hour to generate one minute of top-level tokens plus around 8 hours to upsample one minute of top-level tokens.[2](https://arxiv.org/pdf/2005.0341) The two sources agree on the total but differ slightly in the split.[1](https://openai.com/index/jukebox/) • [2](https://arxiv.org/pdf/2005.0341)
Interest persisted after OpenAI stopped maintaining the project. The repository has approximately 8,029 stars and 1,441 forks,[3](https://github.com/openai/jukebox/) and the architecture was added to [Hugging Face Transformers](https://www.edgepedia.org/hugging-face-transformers) on 2023-06-20 (the model itself was released on 2020-04-30), bringing it into a mainstream library.[6](https://huggingface.co/docs/transformers/main/model_doc/jukebox) That integration supports inference only, citing the extreme memory requirements of training, and [Hugging Face](https://www.edgepedia.org/hugging-face)'s documentation notes that generating one minute of audio with the 5b top prior takes about 8 hours on a V100 GPU.[6](https://huggingface.co/docs/transformers/main/model_doc/jukebox)
## References
1. [Jukebox | OpenAI (announcement blog, April 2020)](https://openai.com/index/jukebox/)
2. [Jukebox: A Generative Model for Music (arXiv:2005.0341)](https://arxiv.org/pdf/2005.0341)
3. [openai/jukebox (GitHub repository)](https://github.com/openai/jukebox/)
4. [Jukebox Sample Explorer](https://jukebox.openai.com/?song=802878241)
5. [openai/jukebox README](https://github.com/openai/jukebox/blob/master/README.md)
6. [Jukebox · Hugging Face documentation](https://huggingface.co/docs/transformers/main/model_doc/jukebox)
---
*Topic: Encyclopedia › Technology and the built world › Computing and digital systems › Modern AI: foundation models, generative AI, and the AI industry › Model families and named models › Audio, music, and speech models*
*Initially written Sep 17, 2026 · Edited: Oct 11, 2026; Oct 15, 2026*
*Copyright 2026 EdgeChat AI, a subsidiary of Biostate AI.*
License: Edgepedia Community License 1.0, https://www.edgepedia.org/edgepedia/license