TL;DR
This paper compares two acoustic methods for measuring speech rhythm, finding that while both can distinguish between Chinese and English, the Grabe & Low (GL) metric offers better graphical separation of individual speakers and genres. The study emphasizes the importance of large, naturalistic datasets and highlights how speech genre (reporter vs. interviewee) significantly impacts rhythm measurements.
What Problem Does This Paper Address?
For decades, linguists have debated whether languages can be categorized by their speech rhythm, broadly into "stress-timed" (like English) and "syllable-timed" (like Spanish). While perceptual differences are clear, objectively measuring these distinctions acoustically has been challenging. Early attempts relied on subjective perception, but more recently, researchers have developed acoustic metrics to quantify speech timing patterns. This paper addresses the question of which of two prominent acoustic measurement approaches – the Ramus, Nespor & Mehler (%V vs. AC) method and the Grabe & Low (PVI) method – is more effective in distinguishing speech rhythm across languages, particularly when using naturalistic, broadcast news speech data.
Methods
The study compares two primary acoustic measurement approaches:
- Ramus, Nespor & Mehler (%V vs. AC): This method plots the percentage of vowel duration (%V) against the standard deviation of consonant duration (AC) within a sentence.
- Grabe & Low (PVI): This approach uses the Pairwise Variability Index (PVI) to measure the variation in durations between consecutive units. It employs a "raw" PVI (rPVI) for intervocalic segments and a "normalized" PVI (nPVI) for vocalic segments, which aims to control for speech rate.
The researchers applied these metrics to a large corpus of broadcast news speech from American English and Mandarin Chinese. They then analyzed the data by averaging measurements across different divisions: language, gender, and genre (reporter vs. interviewee). A key methodological difference from prior work is calculating averages based on the mean of each individual speaker's sub-dataset, rather than averaging all data points within a division, to prevent single speakers from disproportionately influencing the results.
Datasets
The study utilized a large corpus of audio files from American English and Mandarin Chinese broadcast news. This data was originally processed by automatic speech recognition (ASR) systems to produce time-aligned phonetic transcriptions, including phoneme durations. The corpus comprised data from 92 speakers (42 English, 50 Chinese) and approximately 99 "phoneme" minutes of speech after filtering out pauses and commercials. This dataset is the same as used in a previous study ([7]) and is described as "naturalistic" broadcast news, acknowledging that it might represent a "read" speech variety due to teleprompter use, but still considered prototypical for each language.
Key Insights and Ideas
- Genre Matters: Speech genre (reporter vs. interviewee) had a significant impact on rhythm measurements, with reporters clustering more tightly for both languages and metrics. This suggests that "target" rhythms might be more closely approximated by speakers in more formal, broadcast roles.
- Individual Variation is High: While acoustic metrics can statistically distinguish between Chinese and English, individual speakers within each language show considerable overlap, regardless of the metric used. This underscores the need to analyze multiple speakers to draw robust conclusions about language-level rhythm.
- PVI's Graphical Advantage: The Grabe & Low (PVI) metric, particularly its graphical representation, appears to offer better visual separation of individual speakers and genre divisions compared to the RNM (%V vs. AC) method.
- Natural Data Yields Consistent Results: Using a large dataset of naturalistic speech with a significant number of speakers yields similar basic results for both measurement approaches, supporting the validity of these metrics for rhythm classification.
Key Results
- Both the RNM (%V vs. AC) and GL (PVI) metrics indicated statistically significant differences between American English and Mandarin Chinese speech rhythm.
- Table 1 shows that English speakers had higher %V, AV, and AC values than Chinese speakers, contrary to some previous findings.
- Figure 1 (GL metric) showed more overlap between language, gender, and genre divisions than Figure 2 (RNM metric), suggesting the RNM metric provided clearer overall language separation.
- However, Figures 3 and 4 (individual speakers) and Figures 5 and 6 (genre divisions) suggest the GL metric provided better graphical separation, particularly for genre divisions (Figure 5 vs. Figure 6).
- Reporters (REP) in both Chinese and English showed lower standard deviations (SD) in their measurements compared to interviewees (INT) across all tested values (Table 2), indicating tighter rhythm patterns.
Limitations and Future Work
- Speech Rate Normalization: The paper questions whether the normalization in the vocalic PVI (nPVI) effectively controls for speech rate, suggesting further investigation is needed.
- Data Representativeness: While the study used naturalistic broadcast news, the authors acknowledge that the "interviewee" (INT) category might not be as clearly defined as the "reporter" (REP) category. Additionally, the density of speakers within certain categories (e.g., CH INT vs. CH REP) was uneven, potentially affecting the interpretation of density in graphical representations.
- Future Directions: The authors suggest that more data from conversational English and interview-based Chinese speech is needed to see if the PVI can consistently split languages without overlap. They also point to future work exploring rhythm "tendencies" with metrics like the variation coefficient (VarcoV, VarcoC) and sonority, especially in light of studies suggesting rhythm distinctions can be found even without explicit speech rate correction.
Summary updated