Free Music, All the time, Everywhere 🎶
So essentially,
Stable Audio Open makes ethical sourced music for free!
Paper:
Stable Audio Open (6 Pages)
Github:
https://stability-ai.github.io/stable-audio-open-demo/
Researchers from Stability AI are interested in developing text to music models.
Hmm..What’s the background?
Many existing models are either private or lack public weights, hindering further research and artistic applications. The licensing of audio data used in training public models is often not fully documented, raising concerns about copyright infringement. Current open models struggle to match the quality, coherence, and inference speed of state-of-the-art private models.
To address these challenges, the researchers developed a text-conditioned generative model for non-speech audio.
Ok, So what is proposed in the research paper?
The paper has the following key proposals:
This model is intended to promote open research and foster further development and artistic exploration
By relying solely on CC-licensed data, the researchers prioritize ethical and transparent data practices, addressing concerns about copyright infringement prevalent in other text-to-audio models
The researchers aim for their model to produce high-quality audio, competitive with existing private models. They highlight the reported FDopenl results, a measure of realism in generated audio, as evidence of achieving this goal
What’s next?
The researchers share potential areas for improvement and further exploration:
Stable Audio Open's performance in music generation, while surpassing that of the best open model, falls short compared to non-open models like Stable Audio, this can be improved
Fix prompting difficulties in generating audio from prompts containing connectors (e.g., "and", "followed by") or those involving intelligible speech
Stable Audio Open was primarily trained on English text, which may limit its performance in other languages
So essentially,
Stable Audio Open makes ethical sourced music for free!