Research PaperSequential stream segregation of voiced and unvoiced speech sounds based on fundamental frequency

- The stimuli used in the study consisted in an unvoiced fricative consonant and a voiced vowel (CV token).
- Listeners could use a difference in F0 to segregate alternating CV tokens.
- Evidence for both obligatory and voluntary stream segregation was found.
- Listeners did not base their judgments on the vowel part only or the consonant part only.
- Listeners were no longer able to perform the task without the fricative part of the stimuli.
- Listeners were able to segregate the whole tokens based on ΔF0 despite the lack of F0 cues in the fricative part.

Differences in fundamental frequency (F0) between voiced sounds are known to be a strong cue for stream segregation. However, speech consists of both voiced and unvoiced sounds, and less is known about whether and how the unvoiced portions are segregated. This study measured listeners' ability to integrate or segregate sequences of consonant-vowel tokens, comprising a voiceless fricative and a vowel, as a function of the F0 difference between interleaved sequences of tokens. A performance-based measure was used, in which listeners detected the presence of a repeated token either within one sequence or between the two sequences (measures of voluntary and obligatory streaming, respectively). The results showed a systematic increase of voluntary stream segregation as the F0 difference between the two interleaved sequences increased from 0 to 13 semitones, suggesting that F0 differences allowed listeners to segregate speech sounds, including the unvoiced portions. In contrast to the consistent effects of voluntary streaming, the trend towards obligatory stream segregation at large F0 differences failed to reach significance. Listeners were no longer able to perform the voluntary-streaming task reliably when the unvoiced portions were removed from the stimuli, suggesting that the unvoiced portions were used and correctly segregated in the original task. The results demonstrate that streaming based on F0 differences occurs for natural speech sounds, and that the unvoiced portions are correctly assigned to the corresponding voiced portions.