TL;DR
This paper investigates how Japanese listeners distinguish between long and short vowels, finding that while vowel duration is the primary cue, changes in fundamental frequency (F0) within the vowel become crucial when duration alone is ambiguous. This suggests a more nuanced perceptual system for vowel quantity than previously assumed, especially in natural speech where duration can be affected by speaking rate.
What Problem Does This Paper Address?
The paper tackles the question of how listeners identify vowel quantity (the distinction between long and short vowels) in Japanese. While vowel duration is generally accepted as a primary cue for this distinction across languages, its reliability can be compromised in spontaneous speech due to factors like speaking rate. For instance, a vowel that is phonemically short might be lengthened in slow speech, and a phonemically long vowel might be shortened in fast speech, blurring the duration-based distinction. This raises the question: when vowel duration is ambiguous, what other acoustic cues do listeners rely on to identify vowel quantity? Previous research on Japanese suggested spectral cues were not used, leaving fundamental frequency (F0) as a potential candidate. This study aims to systematically investigate the roles of both vowel duration and F0 contour in Japanese vowel quantity perception.
Methods
The researchers conducted a perception experiment using synthesized speech.
- Production: They recorded a native Japanese speaker producing target words with five distinct vowel pairs ([a-a:, e-e:, i-i:, o-o:, u-u:]) embedded in carrier sentences. They identified key acoustic features (duration and F0) in these natural productions.
- Resynthesis: Using the Praat software, they created a 10x10 synthesis matrix for each vowel pair. This matrix systematically varied vowel duration (10 steps) and F0 contour (10 steps). The F0 contour adjustment involved changing the slope within a specific segment of the vowel, moving from a fully descending contour (typical of long vowels) to a flat contour (typical of short vowels). The extreme points of this matrix were based on the selected "natural" productions.
- Perception Experiment: Thirty-one native Japanese listeners participated. They were presented with synthesized words and asked to identify whether the vowel was long or short by choosing from two written options. Reaction times were also recorded.
Datasets
The study used synthesized speech materials created from recordings of a single native Japanese male speaker. The specific dataset consists of 10 synthesized vowel pairs ([a-a:, e-e:, i-i:, o-o:, u-u:]) each presented across 100 synthesized steps (10 duration steps x 10 F0 contour steps). No pre-existing datasets were used for training or evaluation; the synthesized stimuli served as the evaluation material.
Key Insights and Ideas
- Duration Dominance: Vowel duration is confirmed as the primary perceptual cue for Japanese vowel quantity. As synthesized duration increased, the percentage of "long" responses consistently rose.
- F0 as a Backup Cue: When vowel duration alone is insufficient to cue quantity (specifically, around the 50% crossover point where listeners are equally likely to perceive it as long or short), the F0 contour plays a significant role.
- F0 Contour's Role: A descending F0 contour within the vowel tends to be perceived as indicative of a long vowel, while a flat F0 contour is perceived as indicative of a short vowel, particularly when duration is ambiguous.
- Vowel-Specific Differences: The paper notes some differences across vowel types, particularly for /a/-/a:/, where the reliance on F0 might be less pronounced or manifest differently, potentially due to artifacts in the selected production of the long /a:/ vowel.
- Implications for Natural Speech: The findings are crucial for understanding speech perception in natural, spontaneous speech, where duration can be highly variable. The F0 contour acts as a valuable supplementary cue, making vowel quantity identification more robust.
Key Results
- For all five vowel sets, increasing synthesized vowel duration led to a statistically significant increase in the percentage of "long" responses (p < .0001 for all).
- At duration steps where ambiguity existed (the 50% crossover point), listeners' perception shifted significantly based on the F0 contour. Specifically, a flat F0 contour (step 1) led to more "short" responses, while a descending F0 contour (step 10) led to more "long" responses.
- The study found that for duration steps where quantity was clearly cued (extreme duration steps 1 and 10), listeners' responses were largely independent of the F0 contour adjustments.
Limitations and Future Work
- Single Speaker: The study used only one male speaker for production. This limits the generalizability of the findings to other speakers, voices, or accents.
- Synthesized Stimuli: While systematic, synthesized speech may not perfectly replicate the natural acoustic variations present in spontaneous speech.
- Vowel-Specific Anomalies: The paper notes that the /a/-/a:/ pair showed some deviations from the general pattern, suggesting that the interaction between duration and F0 might be vowel-specific or influenced by the specific acoustic properties of the selected productions.
- Future Work: The authors suggest further investigation into the vowel-specific differences and the role of F0 in more naturalistic speech contexts. Exploring how other acoustic features might interact with duration and F0, and examining these cues in different speaking rates or dialects, would be valuable extensions.
Summary updated