TL;DR
This paper investigates the acoustic differences between two distinct ways creaky voice manifests in American English: /t/ glottalization (replacing a /t/ sound with a glottal stop) and phrasal creak (prolonged creaky voice at the end of phrases). The study finds that while both phenomena involve creaky voice, they are acoustically distinguishable, with glottalization characterized by rapid changes in voice noise over time, and phrasal creak by lower pitch and spectral tilt.
What Problem Does This Paper Address?
Creaky voice, a voice quality characterized by a low and irregular fundamental frequency (F0) and increased vocal fold constriction, is not a monolithic phenomenon. In American English, it arises from at least two distinct linguistic processes: /t/ glottalization, where a syllable-final /t/ is realized as a glottal stop (e.g., "about" as [əba?]), and phrasal creak, which involves prolonged creaky voice often occurring at the boundaries of prosodic phrases. While previous research has shown that listeners can differentiate between these two, suggesting acoustic differences exist, the specific acoustic features that enable this distinction, especially in spontaneous speech and across multiple speakers, remained unclear. This study aims to identify these acoustic correlates and understand how they contribute to distinguishing between these two sources of creaky voice.
Methods
The researchers employed a linear discriminant analysis (LDA) to classify tokens of speech into four categories: non-glottalized/non-creaky, non-glottalized/creaky, glottalized/non-creaky, and glottalized/creaky. They extracted eleven acoustic measures, including fundamental frequency (F0), noise measures (Cepstral Peak Prominence - CPP, Harmonics-to-Noise Ratio - HNR05, Subharmonics-to-Harmonics Ratio - SHR), and spectral tilt measures. These measures were analyzed both as averages over the vocalic portion (or vocalic portion plus glottalization) and as changes from the first to the final third of the vowel. To account for speaker-specific variations, all acoustic measures were standardized within each speaker. The LDA was used to find linear combinations of these acoustic features that best separated the groups.
Datasets
The study utilized the Buckeye Corpus of spontaneous American English, a dataset containing recordings of 40 speakers. From this corpus, they selected tokens of syllable-final /t/ that were realized as either a voiceless stop [t] or a glottal stop [?]. These tokens were further filtered based on phonetic criteria (e.g., preceding vowel duration, absence of complex codas) and then hand-inspected for accuracy. Phrasal creak was identified based on corpus annotations and hand-inspection, defined as irregular voicing lasting at least twice the duration of the preceding vowel.
Key Insights and Ideas
- Distinct Acoustic Signatures: The study successfully demonstrates that /t/ glottalization and phrasal creak, despite both involving creaky voice, have distinct acoustic profiles. This supports the idea that different linguistic functions can be realized through acoustically contrastive voice qualities.
- Glottalization: A Dynamic Process: Glottalization is primarily distinguished by changes in acoustic measures over time, particularly a rapid increase in noise (indicated by a decrease in Cepstral Peak Prominence) towards the end of the vowel or glottalization. This suggests that the dynamic transition into the glottal stop is a key acoustic cue.
- Phrasal Creak: A Static Quality: Phrasal creak, on the other hand, is characterized by more static acoustic properties, specifically lower F0 and spectral tilt, reflecting a generally more constricted and irregular vocal fold vibration throughout the affected portion of speech.
- Listener Sensitivity: The findings align with previous work suggesting listeners can differentiate these phenomena, implying that these acoustic differences are perceptually relevant.
- Orthogonal Dimensions: The LDA identified two main discriminant functions, one primarily distinguishing glottalization from non-glottalization (LD1) and the other distinguishing creaky from non-creaky voice (LD2). This suggests these two linguistic processes operate along relatively independent acoustic dimensions.
Key Results
- The LDA successfully classified the four groups above chance, with the first discriminant function (LD1) explaining 86.3% of the variance, primarily distinguishing glottalization. The second discriminant function (LD2) explained 11.4% and primarily distinguished phrasal creak.
- Cepstral Peak Prominence (CPP), particularly its change over time (ACPP), was the most important predictor for distinguishing glottalization in LD1.
- Lower F0 and spectral tilt measures (e.g., H1-H2) were strongly associated with phrasal creak in LD2.
- While glottalized tokens showed a rapid change in noise, creaky tokens (both glottalized and non-glottalized) generally exhibited lower F0 and spectral tilt compared to non-creaky tokens.
- Classification accuracy for glottalization was reasonably successful, but was influenced by prior probabilities. Phrasal creak was also identified, though the study notes that distinguishing creaky from non-creaky tokens was less robust, possibly due to the sample size.
Limitations and Future Work
- Sample Size for Phrasal Creak: The authors note that the relatively smaller number of tokens with phrasal creak in their sample may have impacted the robustness of distinguishing creaky from non-creaky voice.
- Listener Perception: While the study infers perceptual relevance from previous work, it does not directly test listener perception of the identified acoustic features. Future work could investigate whether listeners indeed rely on these specific acoustic cues (e.g., rapid noise changes for glottalization, spectral tilt for phrasal creak) to distinguish between these phenomena.
- Idiosyncrasy: The study used spontaneous speech from multiple speakers, aiming for robustness. However, it acknowledges that previous studies using single speakers might have captured idiosyncratic features. Further research could explore the extent to which these acoustic differences are speaker-dependent.
- SHR Measure: The study found that the Subharmonics-to-Harmonics Ratio (SHR) was not a strong contributor to discrimination, despite its association with multiply-pulsed phonation. Future work could explore other measures of vocal fold vibration patterns.
Summary updated