Overview

AudioGen is part of AudioCraft and turns text descriptions into audio samples that can be saved as WAV files. Its released 1.5-billion-parameter model combines EnCodec audio tokenization with an autoregressive Transformer; the provided reimplementation is a single-stage model trained over a 16 kHz tokenizer with four codebooks sampled at 50 Hz.

The model supports text-conditioned and unconditional generation, continuation from an audio prompt, and several sampling strategies. The example API is configured for five-second samples. The available pretrained model is facebook/audiogen-medium; it is not the original model used to report results in the AudioGen publication.

Key features

Generation and control

Greedy, temperature, top-K, and top-P sampling give developers ways to shape generation behavior. Prompt continuation is useful when a new sound should build from existing audio rather than start from a text description alone. These controls make AudioGen more adaptable than a single fixed text-to-sound path, though satisfying results may require prompt engineering.

Local model and training path

A local Jupyter notebook demo provides a route to trying the model on a GPU. AudioGenSolver implements a training pipeline, but the training datasets are not provided, and the maker cautions that the pipeline may not fully reproduce the paper's results. Teams seeking a repeatable training setup should account for that gap.

Pricing

AudioGen (self-hosted): 0.00 USD per free (billed No price stated by maker). The plan includes the pretrained medium model; local inference requires a GPU with at least 16 GB of memory. There is no free trial, and the model is free to use, not commercially unrestricted: the code is MIT-licensed, while the weights use CC-BY-NC 4.0. That distinction makes AudioGen a poor fit for commercial deployment unless the licensing constraints are resolved.

Platforms

AudioGen is available through an API and for Linux and self-hosted use. Installation requires Python 3.9 and PyTorch 2.1.0; ffmpeg is recommended. The GPU memory requirement makes local inference a meaningful infrastructure commitment, rather than a lightweight option for any developer machine.

Who it's for

Audio and machine-learning researchers, generative-model learners, and teams comfortable managing their own inference stack are the clearest fit. It is less suitable for projects requiring realistic vocals, strong performance across languages, or commercially usable model weights: AudioGen was trained on English descriptions and performs less well in other languages, and it cannot generate realistic vocals. The model card also advises evaluating and mitigating risks before downstream use.

Pros and cons

Pros: Free access to a pretrained model, API support, and self-hosting give technical teams control over their generation workflow without a subscription price.
Pros: Conditional and unconditional generation, prompt continuation, and four sampling approaches offer meaningful options for shaping output.
Cons: The medium model needs a GPU with at least 16 GB of memory, excluding many low-resource local setups.
Cons: The CC-BY-NC 4.0 weight license restricts commercial use even though the code is MIT-licensed.
Cons: English-centered training, weak realistic-vocal generation, and the need for prompt engineering limit the range of projects it serves well.
Cons: Missing training datasets and a pipeline that may not reproduce publication results complicate independent reproduction.

Alternatives

For a broader comparison of sound-effect tools, browse AI Sound Effect Generators; for music-focused tools, see AI Music Generators.

Mirelo is a better fit for teams wanting API, web, desktop, or macOS access and a paid Creator plan with monthly credits and projects.
Async offers API, iOS, and web access; choose it when a media-upload and export workflow matters more than local model control.
Pixazo AI SFX Generator is an option for web-based use with a seven-day credit trial, though its free trial has no API access or commercial rights.
SFX Studio suits users seeking a web/API option with free signup credits and a paid monthly tier.
Sonilo is worth considering for users who want Android or iOS access alongside web and API, with recurring free credits.
Sony Woosh is another free option with API, self-hosted, and web platforms.
Resemble AI is a paid alternative with API, self-hosted, and web platforms.
Morphic may suit web-only users who want a free tier with limited credits and a single user.

Verdict

Choose AudioGen if you need a free, controllable text-to-sound model and can provide GPU-backed inference while working within a non-commercial license. Look elsewhere when commercial rights, realistic vocals, multilingual prompts, or a lighter deployment footprint are essential.

	AudioGen	MakeSFX
Free plan	Yes	Yes
Free trial	No	No
Paid from	Free	Free
Open source	No	No
Platforms	api, Linux, self-hosted	Web
Free plan	Yes	Yes
Text-to-sound	Yes	Yes
Reference input	Yes	—
Download formats	wav	—
Commercial use	restricted	included
API access	Yes	—
Maximum clip length	—	30 s

AudioGen vs MakeSFX

AudioGen

MakeSFX