TL;DR
This paper challenges the traditional linguistic classification of languages into "stress-timed," "syllable-timed," and "mora-timed" based on empirical evidence. It proposes that current rhythm metrics might be more sensitive to the underlying text structure than the actual speech signal, suggesting a need for more sophisticated models of prosodic realization.
What Problem Does This Paper Address?
For decades, linguists have attempted to classify languages based on their rhythmic properties, broadly categorizing them as stress-timed (like English), syllable-timed (like French), or mora-timed (like Japanese). The assumption was that these categories reflected distinct patterns of timing in speech, with stress-timed languages having equally spaced stressed syllables and syllable-timed languages having syllables of roughly equal duration. However, empirical investigations have consistently failed to find robust evidence supporting these distinct categories. This paper addresses the gap between these traditional typological classifications and the actual acoustic realization of rhythm in speech, questioning the validity of existing metrics and proposing a shift towards more nuanced modeling.
Methods
The paper reviews and critiques existing methods for measuring speech rhythm. It discusses the "Ramus metrics," which involve segmenting speech into vocalic and consonantal intervals and calculating parameters like the percentage of vocalic duration (%V) and the standard deviation of consonantal interval durations (ΔC). The authors then demonstrate that similar rhythmic distinctions can be achieved by simply counting phonemes within these intervals, suggesting that these metrics might be capturing textual phonological structure rather than purely acoustic rhythm. They also introduce their own analysis using the Aix-Marsec corpus, employing "Narrow Rhythm Units" (NRUs) and analyzing phoneme durations, z-scores, and the effect of phoneme position within the NRU.
Datasets
- Aix-Marsec corpus: A large corpus of spoken British English (5.5 hours) used for detailed analysis.
- Eurom corpus: Used for replicating Ramus' metrics on comparable sentences read by English, French, and Japanese speakers.
- Data from previous studies: The paper references and analyzes data from various previous studies on English, French, Spanish, Italian, Greek, Dutch, Polish, and Japanese.
Key Insights and Ideas
- Typological classification is empirically weak: The traditional stress-timed vs. syllable-timed distinction is not well supported by acoustic measurements.
- Metrics might reflect text, not just speech: The Ramus metrics, while seemingly robust, may be capturing the underlying phonological structure of the text (e.g., consonant-vowel ratios) rather than the acoustic rhythm of the utterance itself.
- Phoneme counting can mimic acoustic metrics: Simply counting phonemes within vocalic and consonantal intervals can produce similar distinctions between languages as acoustic duration-based metrics.
- Textual structure influences perceived rhythm: Factors like the number of phonemes in a unit and the position of a phoneme within a unit (initial, medial, final) significantly impact duration, potentially more than stress.
- Need for sophisticated models: A simple isochronous stress-foot hypothesis is insufficient. More complex models are needed to account for the interplay of textual structure, prosodic units, and acoustic realization.
Key Results
- Ramus' metrics (%V and ΔC) showed a reasonable separation of English, French, and Japanese, aligning with the traditional rhythmic classes.
- However, using phoneme counts instead of durations yielded similar separation, suggesting the metrics were sensitive to textual phonological structure.
- Analysis of the Aix-Marsec corpus revealed a strong linear correlation between the mean duration of stress feet and the number of syllables within them.
- Mean durations of stressed and unstressed syllables decreased non-linearly with the number of syllables in a foot.
- Analysis of Narrow Rhythm Units (NRUs) showed a significant effect of phoneme position within the NRU, with initial and final phonemes exhibiting different durations.
Limitations and Future Work
- Limitations of current metrics: The paper highlights that current metrics may not fully capture the acoustic realization of rhythm and might be overly influenced by textual properties. The "Klatt's unsolved problem" regarding the appropriate unit for speech timing is reiterated.
- Future Work: The authors plan to explore the interactions between NRU initial/final lengthening, quantal lengthening, and final lengthening at the intonation unit level. They also aim to investigate the applicability of their proposed model of prosodic structure to other languages, particularly French, by analyzing comparable data.
Summary updated