The role of glottal pulse rate and vocal tract length in the perception of speaker identity

Etienne Gaudrain, Su Li, Vin Shen Ban, Roy D. PattersonInterspeech 2009

TL;DR

This paper investigates how listeners perceive speaker identity based on two key vocal characteristics: glottal pulse rate (GPR) and vocal tract length (VTL). They found that while listeners are more sensitive to changes in GPR, they rely more heavily on VTL for identifying a speaker, suggesting that our perception of speaker identity is shaped by our expectations of how these vocal parameters vary within a single speaker's speech.

What Problem Does This Paper Address?

Identifying a speaker from their voice is a complex process that relies on various acoustic cues. Two prominent cues are the glottal pulse rate (GPR), which relates to pitch, and the vocal tract length (VTL), which influences the resonant frequencies of the vocal tract and thus the timbre of the voice. Previous research has shown that both GPR and VTL play a role in perceived speaker size and sex. However, the relative importance of these cues for speaker identity perception, and how listeners' expectations about natural vocal variation influence this perception, remained unclear. Specifically, while listeners can discriminate small changes in GPR, natural speech exhibits much larger variations in GPR than in VTL for a single speaker. This raises the question: does this natural variation in variability lead listeners to prioritize VTL for identity judgments, even if they are more sensitive to GPR changes?

Methods

The researchers created synthetic speech stimuli using the STRAIGHT vocoder to manipulate GPR and VTL independently. They generated five "reference speakers" with distinct combinations of GPR and VTL, representing categories like "Man," "Woman," "Child," "Dwarf," and "Castrato." Each reference speaker was then surrounded by 48 "comparison speakers" whose GPR and VTL were systematically varied along eight radial spokes in the GPR-VTL plane. Participants listened to pairs of short CV syllable sequences (one reference, one comparison) and were asked to judge whether they could have been uttered by the same speaker. This "similarity judgment" task differs from simple discrimination tasks, aiming to capture how listeners form expectations about speaker identity. Statistical analysis involved a linear mixed-effects model to assess the influence of radial distance, spoke, and reference speaker on similarity judgments. They also explored a "trading relationship" between GPR and VTL, investigating how much change in one parameter is equivalent to a change in the other for maintaining perceived similarity.

Datasets

The study did not use pre-existing datasets for training. Instead, they generated their own synthetic stimuli. The core "dataset" consists of 2450 unique comparisons per subject, formed by pairing five reference speakers with 48 comparison speakers each, repeated ten times. The stimuli were 65 CV syllables spoken by a single original speaker, processed to have a duration of 250 ms.

Key Insights and Ideas

  • Experience Trumps Sensitivity: Listeners are more sensitive to small changes in GPR (lower just-noticeable difference), but they require a much larger change in GPR (3.8 semitones) than in VTL (2.2 semitones) to perceive a speaker change. This suggests that our perception of speaker identity is driven more by our learned expectations of natural vocal variation than by raw perceptual sensitivity.
  • GPR Variability is Key: The natural speech observation that GPR varies much more within a single speaker than VTL appears to be a crucial factor. Listeners seem to use this knowledge to infer speaker identity, accepting larger GPR shifts as "normal" for a given speaker.
  • Inverse Trading Relationship: The study found an inverse trading relationship compared to previous work on speech segregation or discrimination. For speaker identity, a change in GPR was equivalent to a smaller change in VTL (GPR change of 1 semitone ≈ VTL change of 0.6 semitones), whereas for discrimination tasks, VTL changes needed to be larger.
  • Boundary Effects: Listener confidence in their judgments decreased near the boundaries of the GPR-VTL plane, suggesting that extreme vocal characteristics might be harder to categorize or lead to less certain identity perceptions.

Key Results

  • Listeners judged two sequences as being from the same speaker if the VTL difference was approximately 2.2 semitones, but required a GPR difference of about 3.8 semitones.
  • The GPR-VTL trading ratio for speaker identity judgment was found to be approximately 0.60, meaning a 1 semitone change in GPR was equivalent to a 0.6 semitone change in VTL.
  • Statistical analysis revealed significant effects of radial distance, spoke, and reference speaker on similarity judgments, along with significant interactions between these factors.

Limitations and Future Work

  • Stimulus Generation: The study used synthetic stimuli, which, while allowing precise control, may not perfectly capture the nuances of natural speech. The authors acknowledge that the "fidelity of stimulus manipulation" could influence judgments, especially at the boundaries of the GPR-VTL plane.
  • Limited Speaker Categories: While five reference speaker categories were used, the full spectrum of human vocal variation is much broader.
  • Future Work: The authors suggest further investigation into how experience with different types of speech (e.g., emotional speech, different languages) might shape these expectations and influence speaker identity perception. Exploring the precise mechanisms behind the boundary effects and how they relate to categorical perception of voices would also be valuable.

Summary updated

BibTeX
@inproceedings{gaudrain2009role,
  author = {Etienne Gaudrain and Su Li and Vin Shen Ban and Roy D. Patterson},
  title = {{The role of glottal pulse rate and vocal tract length in the perception of speaker identity}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure