AudioGen vs EzAudio
AudioGen
github.comEzAudio
haidog-yaqub.github.ioSide by side
| Figure | AudioGen | EzAudio |
|---|---|---|
| Price from | Free | — |
| Free plan | Yes | — |
| Free trial | — | — |
| Open source | — | — |
| Web | — | Yes |
| Linux | Yes | — |
| API | Yes | — |
| Self-hosted | Yes | Yes |
| Plans published | 1 | 0 |
| Spec score · AI Sound Effect Generators | 8.8 (#2) | 7.3 (#22) |
| License | The repository says its code is MIT-licensed and its model weights use CC-BY-NC 4.0. | The GitHub repository identifies its license as MIT. |
| Model design | The provided reimplementation is a single-stage autoregressive Transformer trained over a 16 kHz EnCodec tokenizer with four codebooks sampled at 50 Hz. | EzAudio uses an optimized diffusion transformer for audio latent representations and a one-dimensional waveform variational autoencoder. |
| Purpose | AudioGen is a text-to-sound generation model provided through AudioCraft. | EzAudio is a diffusion based model that generates audio from text prompts. |
| Training | AudioGenSolver implements the training pipeline, but the maker says it may not fully reproduce the paper results and does not provide the AudioGen training datasets. | The README gives a training command for the text-to-audio diffusion model and points to the SoloAudio work for autoencoder training. |
Plans
In their own words
AudioGen
AudioGen is a text-to-sound model provided through AudioCraft. It generates audio samples from text descriptions and can continue from an audio prompt, with conditional and unconditional generation modes. Sampling controls include greedy, temperature, top-K, and top-P methods. The listed pretrained model, facebook/audiogen-medium, has 1.5 billion parameters; the example API generates sound and saves it as WAV. The implementation uses an autoregressive Transformer and a 16 kHz EnCodec tokenizer. The repository provides API access and a local Jupyter notebook demo. Running the medium model locally requires a GPU with at least 16 GB of memory; installation requires Python 3.9 and PyTorch 2.1.0, with ffmpeg recommended. The model was trained on English descriptions and performs less well in other languages. It cannot generate realistic vocals, and prompts may need refinement. Code is MIT-licensed, while weights use CC-BY-NC 4.0; training datasets are not provided. The listed self-hosted plan is free.
EzAudio
EzAudio has no write-up on Specifiction yet; its figures are above.