“Recent advancements in neural speech synthesis technologies have brought aboutwidespread applications but have also raised concerns about potential misuse and abuse.Addressing these challenges is crucial, particularly in the realms of forensics and intellec-tual property protection. While previous research on source attribution of synthesizedspeech has its limitations, our study aims to fill these gaps by investigating the identifi-cation of sources in synthesized speech. We focus on analyzing speech synthesis modelfingerprints in generated speech waveforms, emphasizing the roles of the acoustic modeland vocoder. Our research, based on the multi-speaker LibriTTS dataset, reveals twokey insights: (1) both vocoders and acoustic models leave distinct, model-specific fin-gerprints on generated waveforms, and (2) vocoder fingerprints, being more dominant,may obscure those from the acoustic model. These findings underscore the presence ofmodel-specific fingerprints in both components, suggesting their potential significance insource identification applications.”

“To enhance the effectiveness of fake audio detection techniques, researchers have developed mul-tiple datasets such as those for the ASVspoof and ADD challenges. These datasets typically focuson capturing non-emotional characteristics in speech, such as the identity of the speaker and theauthenticity of the content. However, they often overlook changes in the emotional state of theaudio, which is another crucial dimension affecting the authenticity of speech. Therefore, thisstudy reports our progress in developing such an emotion fake audio detection dataset involvingchanging emotion state of the origin audio named EmoFake. The audio samples in EmoFake aregenerated using open-source emotional voice conversion models, intended to simulate potentialemotional tampering scenarios in real-world settings. We conducted a series of benchmark ex-periments on this dataset, and the results show that even advanced fake audio detection modelstrained on the ASVspoof 2019 LA dataset and the ADD 2022 track 3.2 dataset face challengeswith EmoFake. The EmoFake is publicly available1 now.”

“Reinforcement learning with human feedback for aligning large language models (LLMs) trainsa reward model typically using ranking loss with comparison pairs. However, the training pro-cedure suffers from an inherent problem: the uncontrolled scaling of reward scores during rein-forcement learning due to the lack of constraints while training the reward model. This paperproposes a Prior Constraints-based Reward Model (PCRM) training method to mitigate thisproblem. PCRM incorporates prior constraints—specifically, length ratio and cosine similaritybetween outputs of each comparison pair—during reward model training to regulate optimiza-tion magnitude and control score margins. We comprehensively evaluate PCRM by examining itsrank correlation with human preferences and its effectiveness in aligning LLMs via RL. Exper-imental results demonstrate that PCRM significantly improves alignment performance by effec-tively constraining reward score scaling. As another bonus, our method is easily integrated intoarbitrary rank-based alignment methods, such as direct preference optimization, and can yieldconsistent improvement. The code is available at https://github.com/wangclnlp/DeepSpeed-Chat-Extension/tree/PCRM.”