Voice Privacy in Speech Systems: A Comparative Study of Pitch Shifting and StarGAN-VC

Voice recognition systems facilitate naturalistic human−com-puter interaction. However, spoken input may inherently expose sensitive acoustic features that can threaten user privacy. In particular, raw spoken language data can reveal paralinguistic information such as emotional state, health condition, and speaker identity, which poses a significant privacy risk when the speaker’s voice is recognizable, especially within identifiable communities or groups. This study aims to investigate the preservation of acoustic privacy by evaluating two voice transformation techniques: traditional pitch shifting and the StarGAN-VC deep generative model [3], in terms of their effectiveness in obfuscating speaker identity while preserving lexical intelligibility. We measure their performance along two dimensions: lexical accuracy, assessed via an automatic speech recognition (ASR) application programming interface (API), and speaker identifiability, evaluated through subjective human listener studies. Our results show that although both methods degrade ASR performance, StarGAN-VC offers significantly greater privacy protection among individuals within the same social circle, by reducing speaker recognizability with minimal impact on lexical intelligibility. These findings highlight deep generative voice conversion models as viable tools for privacy-preserving solutions in voice-enabled technologies.