I have been thinking of implementing a utilities for watermarking within the speechbrain library. I just wanted to share my thoughts here and open a discussion to see if others agree, or have thoughts of their own.

Watermarking in audio is not new, and recent developments in neural methods are allowing for more robust implementations.
The primary use cases I can see:

1. Responsible development: With the potential for conversational agents being used for purposes to deceive others (e.g. deepfake), it may benefit developers to identify if their agent has been used by bad actors to generate speech for such purposes.

2. Linking modified data to source data: Watermarking files could be utilised as a way to identify a source file. For example say your data goes through some iterations of augmentation or slicing, with a key:value database you could link back to the source file without having to build a graph of relations.

In the first instance, I was thinking of adding an interface to utilise wavmark with default settings, but this could expanded later.

What do people think, would this be a useful addition to the speechbrain library? I'm open to suggestions.

References:

• WavMark: Watermarking for Audio GenerationWavMark: Watermarking for Audio Generation
• WavMark - Github
• A Comprehensive Real-World Assessment of Audio Watermarking Algorithms: Will They Survive Neural Codecs?