Masked-speech Recognition Using Human and Synthetic Cloned Speech

Area of Science:

• Speech processing and human-computer interaction.

Background:

• Voice cloning synthesizes speech mimicking human vocal characteristics.
• Evaluating synthetic speech requires comparison with human speech for intelligibility and human-likeness.

Purpose of the Study:

• To compare intelligibility, human-likeness, and perceptual similarity of human and voice-cloned speech.
• To assess the utility of voice clones in speech recognition research.

Main Methods:

• Masked-sentence recognition tests were conducted with young adults with normal hearing.
• Speech from five human talkers and their voice clones was presented in noise at -6 dB SNR.
• Participants rated human-likeness and identified voice origins (human vs. clone).

Main Results:

• Sentence recognition scores were similar for human and voice-cloned speech, though worse for nonsense than meaningful sentences.
• Human-likeness ratings were comparable between human and cloned voices.
• Participants accurately identified different voices but struggled to distinguish human/clone pairs, often mistaking clones for humans.

Conclusions:

• Voice clones show potential as a reliable tool for evaluating speech recognition, comparable to human speech.
• Voice synthesis and automatic speech recognition are valuable for studying human speech perception.