search

facebookresearch / audioseal

Localized watermarking for AI-generated speech audios, with SOTA on robustness and very fast detector
MIT License
423 stars 51 forks source link

more bit？？？？ #46

Closed DanMerry closed 2 months ago

DanMerry commented 2 months ago

how to embed more bits into the audio?

antoine-tran commented 2 months ago

Hi, could you elaborate ? Did you mean the bits of the secret messages ?

You can modify the existing model card to have for example 32 instead of 16 bit-message in here. For example:


model = AudioSeal.load_generator("audioseal_wm_16bits", nbits=32)
...

Note that this might not always be optimal as the model is trained with different pertubed / augmented training data while embedding 16-bit messages. So other option is to train a new model (see our  instruction notes) , where you can specify the number of bits during the training. We provide an example recipe here, where this line is where the number of bits used for the training is specified

antoine-tran commented 2 months ago

@DanMerry I closed the issue. Feel free to re-open it if you still face the problem