GAMW
A generation-aware Mel-domain audio watermarking method that embeds content-adaptive multi-bit watermark residuals into the intermediate Mel-domain representation during TTS synthesis. It is used for tracing synthesized speech and exhibits strong robustness against compound perturbations.
Timeline (last 90 days)
GAMW (Generation-Aware Mel-Domain Watermarking) is a method that embeds a watermark into the intermediate representation in the mel domain during the TTS synthesis process, rather than post-processing the waveform after synthesis is complete.
GAMW embeds a content-adaptive multi-bit watermark residual into the intermediate representation produced during the TTS generation process.
GAMW's decoding objective function combines two mechanisms: focal modulation and dynamic per-bit weighting.
Under a 16-bit watermark payload, GAMW's decoding accuracy in the single-perturbation scenario is 9.2 percentage points higher than the strongest baseline method.
Under a 16-bit watermark payload, GAMW's decoding accuracy in the composite-perturbation scenario is 23.5 percentage points higher than the strongest baseline method.
The GAMW research team trained and evaluated the method on two different TTS backbones.
GAMW can still maintain reliable clean speech decoding capability under higher test payloads exceeding 16 bits.
GAMW maintains competitive speech quality, meaning watermark embedding does not come at the cost of significantly sacrificing audio quality.
Mel-domain embedding requires the watermarking system to be able to access the internal pipeline of the TTS, which raises the integration barrier for application scenarios that only want to add watermarks at the audio output end.
GAMW has currently only been validated on two TTS backbones, and its applicability under different codecs, transmission links, adversarial attacks, and other conditions in actual deployment still requires more testing.
1 more timeline events
All Facts (11)
GAMW (Generation-Aware Mel-Domain Watermarking) is a method that embeds a watermark into the intermediate representation in the mel domain during the TTS synthesis process, rather than post-processing the waveform after synthesis is complete.
50%UnverifiedGAMW embeds a content-adaptive multi-bit watermark residual into the intermediate representation produced during the TTS generation process.
50%UnverifiedGAMW's decoding objective function combines two mechanisms: focal modulation and dynamic per-bit weighting.
50%UnverifiedUnder a 16-bit watermark payload, GAMW's decoding accuracy in the single-perturbation scenario is 9.2 percentage points higher than the strongest baseline method.
50%UnverifiedUnder a 16-bit watermark payload, GAMW's decoding accuracy in the composite-perturbation scenario is 23.5 percentage points higher than the strongest baseline method.
50%UnverifiedThe GAMW research team trained and evaluated the method on two different TTS backbones.
50%UnverifiedGAMW can still maintain reliable clean speech decoding capability under higher test payloads exceeding 16 bits.
50%UnverifiedGAMW maintains competitive speech quality, meaning watermark embedding does not come at the cost of significantly sacrificing audio quality.
50%UnverifiedMel-domain embedding requires the watermarking system to be able to access the internal pipeline of the TTS, which raises the integration barrier for application scenarios that only want to add watermarks at the audio output end.
50%UnverifiedGAMW has currently only been validated on two TTS backbones, and its applicability under different codecs, transmission links, adversarial attacks, and other conditions in actual deployment still requires more testing.
50%UnverifiedThe code for GAMW has been open-sourced at the GitHub repository https://github.com/ZConJugaTe/GAMW.
50%