Source-linked AI summary
Counting Closures in Spanish Trills: A Multi-Corpus Acoustic Study
Mateo Cámara, Maria F. Alcala-Durand
TL;DR
Large-scale acoustic evidence on Spanish trill structure is scarce, and envelope-peak counters can conflate each closure with its release. The paper applies a closure-based detector with a quality filter and an independent period check to thousands of tokens across six corpora. It finds two closures and a period near 36 ms as reference values for well-formed trills, with context effects but no robust sex effect once closures are counted directly.
Problem
Large-scale acoustic evidence on trill closure count and timing across speakers, contexts, and dialect regions is scarce, while envelope-peak counters can conflate closures with releases.
Source: §1 · p. 1Method
A closure-based detector locates energy minima with verified releases, applies a fixed quality filter, and uses an independent autocorrelation-based period estimator as an internal check.
Source: §1 · p. 1, §4.3 · p. 3, §4.4 · p. 3Results
Takeaways & Limitations
The measured reference values characterize well-formed, voiced, periodic trills, and direct closure counts show no robust evidence of a speaker-sex effect.
Source: §7 · p. 4, §8 · p. 4The study lacks a manually annotated benchmark, and agreement with the cross-detector cannot substitute for one because both estimators use the same envelope.
Source: §7 · p. 4
Abstract
from arXiv · showhide
The Spanish trill /r/ is canonically described as a short sequence of lingual closures, yet large-scale acoustic evidence across corpora is scarce, and automatic counters locating envelope peaks tend to conflate each closure with its release. We present a closure-based detector that locates closures gated by a quality filter and cross-checked against an independent autocorrelation-based period estimator. Applied to 3,560 well-formed (voiced, periodic) trill tokens from 356 speakers across six Spanish corpora, the detector yields a median of two closures and an inter-closure period near 36ms, matching the descriptive literature on all corpora. At the speaker level, phonotactic context is the only factor with a robust, medium effect: onset trills (word-initial and post-/n,l,s/) show more closures than intervocalic rr. We find no robust evidence of a sex effect once closures are counted directly. We report reference values and release a reproducible measurement pipeline for Spanish trills.
1. Introduction
Large-scale acoustic evidence on Spanish trill structure remains fragmented, and envelope-peak counters can conflate closures with releases. This study proposes direct closure counting and applies it across six corpora to examine closure structure and its speaker- and context-level variation.
1. Introduction
- Traditional estimates of two to three contacts come mainly from limited speaker samples, controlled styles, or single varieties, leaving closure count and timing patterns understudied at scale.Sociophonetic work has largely focused on categorical alternations and their conditioning factors.Source: §1 · p. 1
- Envelope-peak counters can count both an occlusion and its release, roughly doubling counts in a way that covaries with voicing spectral density and speaker characteristics.The paper argues for counting closures—occlusion events with a verified release—rather than envelope peaks.Source: §1 · p. 1
- Across six Spanish corpora, the pipeline analyzes 3,560 trill tokens from 356 speakers and reports a median of two closures and an inter-closure period near 36 ms.These values are described as consistent with classic descriptions; the study also examines speaker- and context-level factors.Source: §1 · p. 1
2. Related Work
Prior work describes trill realization as dynamically shaped by articulatory and aerodynamic conditions, while acoustic and sociophonetic findings vary across contexts and varieties. The paper addresses the scarcity of large-scale, closure-resolved corpus studies with a closure-anchored detector and robustness checks.
2. Related Work
- Small changes in pressure or flow can shift lingual trilling toward frication or an approximant, contributing to its cross-linguistic and diachronic instability.Source: §2 · p. 1
- Classic Spanish acoustic work gives baseline ranges for closure structure, voicing, and noise, while treating trillness as gradient and multidimensional.Partially voiced, irregular, or noisy trills make closure structure difficult to estimate at scale.Source: §2 · p. 1
- Reported closure patterns vary by dialect and style: Dominican and Jerezano Andalusian Spanish are modally one-closure, while Venezuelan spontaneous speech is two-closure.Veracruz Spanish can maintain the tap/trill contrast through duration when contacts reduce.Source: §2 · p. 1
- Prior findings on speaker sex conflict, and large-scale closure-resolved corpus studies remain scarce; this study responds with a closure-anchored detector, effect sizes, and cross-dataset robustness checks.Source: §2 · pp. 1–2
3. Data and labeling
The study draws on six Spanish speech corpora spanning Peninsular and American varieties and read and spontaneous speech. It extracts three phonotactically expected trill contexts, then applies a fixed quality filter to define the analyzed set and harmonizes available speaker metadata.
3.1. Corpora
- Six corpora provide phone-level segmentation, spanning Peninsular and American varieties and read and spontaneous speech; three use native alignments and three are force-aligned with a Spanish model.M-AILABS and Common Voice were excluded by a pre-specified policy.Source: §3.1 · p. 2
3.2. Trill token extraction and context labeling
- Tokens come from intervocalic rr, word-initial r, and r after /n, l, s/; tap contexts, consonant clusters, and coda r are excluded.Orthographic constraints and corpus phone labels define the three context categories.Source: §3.2 · p. 2
3.3. Quality filter and analysis set
- The fixed filter retains 3,560 of 16,836 candidates (21%) from 356 speakers, requiring 50–200 ms duration, at least 80% voicing, and periodic modulation score ≥0.40.Retention ranges from 45% in Heroico to 1.3% in DIMEx100; the resulting medians characterize well-formed trills by construction, not the rate of trilling in running speech.Source: §3.3 · p. 2
3.4. Speaker metadata
- Sex comparisons use documented labels for 217 speakers (76 F, 141 M); country and region are not analyzed because the filtered sample is too nationally unbalanced.Source: §3.4 · p. 2
4. Measures and statistical reporting
The measurement pipeline treats verified occlusions as closures, extracts acoustic measures, and detects closures from band-limited energy minima followed by releases. A fixed quality filter and independent period estimator support internal checks, while speaker-level aggregation and effect-size reporting guide inference.
4.1. Terminology
- The study defines a closure as an occlusion with a following release and reserves cycle for the full occlusion–release interval.Closures are the primary count because envelope-peak counters can count both occlusion and release maxima, with bias tied to voicing spectral density.Source: §4.1 · p. 2
4.2. Acoustic measures
- For each token, measures include closure count, duration, mean inter-closure period, closure rate, and voicing percentage; mean f0 is used to test the counting mechanism.Source: §4.2 · p. 2
4.3. Closure detection
- Closure candidates are mid-band envelope minima retained by amplitude criteria and verified by a release 3–25 ms later; a periodicity step can remove outlier intervals when there are at least three closures.The detector uses low- and mid-band RMS envelopes; thresholds were frozen across corpora after synthetic tests and an initial corpus pass.Source: §4.3 · pp. 2–3
4.4. Quality filter
- Tokens pass a fixed filter requiring 50–200 ms duration, at least 80% voicing, and envelope-periodicity score ≥0.40; the score is autocorrelation-based and independent of the closure detector.The score is undefined below 60 ms, making the effective duration floor 60 ms.Source: §4.4 · p. 3
4.5. Independent cross-detector
- A second estimator derives closure count by autocorrelation-based period estimation and duration-to-period rounding, providing an internal consistency check rather than an accuracy test.It shares no closure-event logic with the detector, but both use mid-band modulation and the filter selects periodic tokens.Source: §4.5 · p. 3
4.6. Aggregation and inference
- Speaker-level factors are tested after aggregation to one observation per speaker, while context uses one per speaker×context; tests report effect sizes and bootstrap median intervals.Token-level mixed-effects models also control for context, corpus, duration, and style, with sex added for the documented-sex subset.Source: §4.6 · p. 3
5. Results
Across the analyzed tokens, closure counts match canonical descriptions; phonotactic context is the only factor with a medium effect. Direct closure counts show no robust sex effect, while speaking style has small effects and age and education show no significant effects.
5.1. Closure counts match canonical descriptions
- 2 closures (median) and ∼36 ms (median inter-closure period) characterize 3,560 tokens, with corpus medians of 30–38 ms.The values fall within the descriptive literature’s two-to-three-contact and 30–50 ms ranges.Source: §5.1 · p. 3
5.2. Phonotactic context is the only robust effect
- ε2 = 0.083 for closure count and ε2 = 0.128 for closure rate; onset trills show more closures than intervocalic rr.Intervocalic tokens have the fewest closures and slowest rate; all context medians remain near two closures and overlap substantially.Source: §5.2 · p. 3
5.3. No robust evidence of a sex effect
- 2 closures was the median for both sex groups, with no robust sex effect on any measure.The mixed model estimated a male difference of +0.04 closures (95 % CI [−0.06, 0.13], p = 0.44); envelope-peak counts showed a much larger apparent male advantage.Source: §5.3 · p. 3
5.4. Other factors are weak or null
- ε2 = 0.04 for duration and ε2 = 0.02 for closure rate; age and education showed no significant effects.Read speech was longer and slightly slower than spontaneous speech; age and education results rely on small, uneven bins.Source: §5.4 · p. 3
6. Robustness analyses
Robustness analyses preserve the context and sex findings across corpora and analysis choices. The production results also remain within literature-anchored ranges, while detector agreement is treated as an internal consistency check rather than ground truth.
Per-corpus descriptive checks
- 1–2 closures and 30–38 ms median periods span all six corpora under the production configuration.These values remain within the literature-anchored descriptive ranges.Source: §6 · p. 3
Counted unit and apparent sex difference
- 2 closures but ∼5 envelope peaks appear in one trill; the apparent male advantage is +1.79 envelope peaks versus +0.22 closures, n.s.The figure illustrates why the counted unit changes the sociophonetic conclusion.Source: Figure 1 · p. 4
Per-corpus detector agreement
- 91–96 % of tokens in the four larger corpora show detector agreement within ±1 closure; this is an internal consistency check, not ground truth.The two smallest corpora are noisier, with 67–79 % agreement, but remain within the descriptive ranges.Source: Table 2 · p. 4, §6 · p. 4
Robustness across corpora and analysis choices
- 6/6 corpus hold-outs preserve intervocalic rr as the lowest-count, slowest-rate context; onset effects persist in adjusted models.At fixed duration, onset contexts retain +0.22 and +0.23 closures over intervocalic rr, and the sex null remains unchanged.Source: §6 · p. 4
Pipeline sensitivity analyses
- 15-cell quality-filter sweeps kept the median at two closures, context medium, and sex non-significant.The detector sweep left outcomes unchanged; all six corpora passed for any threshold ≥0.40, supporting the production point of 5 dB and 0.40.Source: §6 · p. 4
Selective attrition check
- 28.1 % versus 29.7 % retention for women and men, respectively, provides no evidence of sex-biased inclusion (p = 0.14).The check covered 8,286 candidates with documented sex, excluding DIMEx100.Source: §6 · p. 4
7. Discussion
Well-formed trills center on two closures and a roughly 36 ms period, while closure-based counts show no robust sex effect. The discussion also identifies a missing expert-annotated benchmark for validating closure detection.
7. Discussion
- Two closures and a ∼36 ms period characterize well-formed, voiced, periodic trills across the corpora.The quality filter excludes reduced variants, so these reference values do not estimate how often trilling succeeds in running conversation.Source: §7 · p. 4
- The context effect persists at fixed duration, and an independent period cross-detector reproduces its ranking.The cross-detector is insensitive to closure depth and, if anything, counts intervocalic trills higher.Source: §7 · p. 4
- Envelope-peak counting yields a large, significant sex effect on the same audio, while closure-based counting does not.Peak counts include both occlusions and releases and covary with f0; anchoring on energy minima with verified releases removes this confound.Source: §7 · p. 4
- The cross-detector cannot replace a manually annotated benchmark because both methods read the same envelope.The authors identify expert closure counts on a balanced subset as future validation work.Source: §7 · p. 4
8. Conclusion
Across six Spanish corpora, well-formed voiced trills center on two closures with a roughly 36 ms period. Phonotactic context is the only robust effect, whereas closure-based counting provides no robust evidence of a sex effect.
8. Conclusion
- Phonotactic context is the only robust effect across the six Spanish corpora.Onset trills—word-initial and post-/n,l,s/—show more closures than intervocalic trills.Source: §8 · p. 4
- Onset trills show more closures than intervocalic trills, but the word-initial estimate rests on 73 tokens.The authors describe that estimate as tentative pending a balanced corpus.Source: §8 · p. 4
- Closure-based counting finds no robust evidence of a speaker-sex effect.The paper cautions that envelope-peak counts covary with f0, so earlier demographic findings may partly reflect the counting method.Source: §8 · p. 4
- Well-formed voiced trills center on two closures with a ∼36 ms period.These reference values are reported across six Spanish corpora.Source: §8 · p. 4