TL;DR
This paper explores how spectral analysis can distinguish between Mandarin and Korean sibilant fricatives, which sound similar but have subtle acoustic differences. The researchers found that specific spectral measures (peak frequency and bandwidth) can effectively differentiate these sounds, even in the speech of Mandarin speakers learning Korean, suggesting a reliance on acoustic similarity rather than strict phonemic categories.
What Problem Does This Paper Address?
The paper tackles the challenge of distinguishing between sibilant fricatives in Mandarin Chinese and Korean. While both languages have sounds that appear similar (e.g., Mandarin's [s], [c], [f] and Korean's [s], [sh], [c], [ch]), their precise articulation and acoustic properties differ. Mandarin has three distinct sibilant places of articulation, with the alveolopalatal [c] often in complementary distribution with the alveolar [s] and postalveolar [f]. Korean, on the other hand, has two primary alveolar fricatives, [s*] and [sh], which contrast in their laryngeal settings (tense vs. breathy voice) and aspiration, leading to spectral differences. Furthermore, both languages have palatalized variants that occur in similar phonetic environments. A key question is whether these subtle acoustic differences can be reliably captured by spectral analysis and how these sounds are perceived and produced by second language learners.
Methods
The researchers employed two spectral measures derived from an auditory-inspired representation:
- Peak ERB Frequency (peakERB): This measure captures the frequency at which the total output power of a gammatone filterbank is highest. It's used to reflect differences in the front cavity size of the vocal tract during fricative production.
- Diffuseness Index (DI): This index quantifies the bandwidth of the spectral peak. It's calculated by measuring the spread of power around the peakERB, specifically looking at the points where the power drops by 9 dB. A higher DI indicates a broader spectral peak.
These measures were applied to 40 ms segments of word-initial fricatives from recorded speech. The spectral representation was generated using a 361-channel, 4th-order gammatone filterbank, with channel center frequencies spaced at 0.1 ERB intervals.
Datasets
The study used recordings from:
- 10 native Mandarin speakers (L1M) who were learning Korean as a second language (L2K).
- 7 native Korean speakers (L1K).
The speech materials consisted of randomized lists of Mandarin and Korean words, each containing 18 target sibilant-initial words. For Mandarin, there were three words each for [sa], [fa], [cja], [si], [fi], and [ci]. For Korean, there were five words each for [sa] and [sha], and four each for [ci] and [chi].
Key Insights and Ideas
- Spectral Measures as Distinguishing Features: The study demonstrates that peakERB and DI are effective in capturing phonemic and allophonic distinctions between Mandarin and Korean sibilants.
- Front Cavity Size Matters: peakERB successfully differentiated between Mandarin [s], [c], and [f], and also between Korean [sh] and [s*], even when they are phonetically similar. This highlights the role of front cavity variations in shaping the acoustic signal.
- Bandwidth Reflects Articulation: The DI measure further refined these distinctions, showing that Mandarin [c] has a broader spectral peak than [s] or [f], and Korean [sha] has a broader peak than [s*a]. This suggests that the degree of constriction or turbulence influences the spectral bandwidth.
- L2 Learners Rely on Acoustic Similarity: A significant finding is that L1 Mandarin speakers learning Korean tend to assimilate Korean sibilants to the acoustically closest Mandarin categories. They don't necessarily maintain the Korean phonemic distinctions, instead mapping them based on spectral similarity.
- Vowel Context Influence: The study observed that the spectral characteristics of sibilants can vary depending on the following vowel, with differences being more pronounced before the low vowel /a/ than the high vowel /i/.
Key Results
- Mandarin Sibilants: In Mandarin, [s] had the highest peakERB, followed by [c], and then [f]. The DI measure showed [c] having a broader peak than [s] or [f].
- Korean Sibilants: Korean [s] generally had a higher peakERB than [sh] before /a/, but this difference was less pronounced before /i/. The DI measure showed [sha] having a broader peak than [sa] before /a/.
- L2 Korean Productions:
- Mandarin speakers learning Korean assimilated both Korean [s*] and [sh] to Mandarin [s] before /a/.
- They assimilated Korean [chi] and [c*i] to Mandarin [c] before /i/.
- In L2 productions, the distinction between Korean [s*] and [sh] was largely lost before /a/, with DI values becoming nearly identical.
Limitations and Future Work
- Vowel Quality Variation: The paper acknowledges that the quality of the high vowel [i] can vary between Mandarin and Korean, which might influence the sibilant-vowel interactions.
- L2 Perception: While the study provides evidence for L2 production patterns, it suggests that further research into L2 perception would be valuable to confirm how Mandarin speakers perceive Korean sibilants.
- Broader L2 Learner Groups: The study focused on L1 Mandarin speakers learning Korean. Future work could explore other L2 learner groups or different language pairs.
- More Complex Phonetic Environments: The analysis focused on word-initial fricatives before /a/ and /i/. Investigating other phonetic contexts could provide a more comprehensive understanding.
Summary updated