TL;DR
This paper investigates how the timing of speech, beyond just the sound quality, can reveal unique characteristics of individual speakers. By analyzing spontaneous speech from a new corpus of Zurich German speakers, they found significant rhythmic differences that could contribute to speaker identification, particularly in vocalic and peak-to-peak interval timings.
What Problem Does This Paper Address?
Traditional speaker identification methods heavily rely on spectral characteristics of speech, which are influenced by the physical anatomy of the vocal tract. While effective to a degree, spectral analysis has limitations. This paper proposes that temporal features of speech – how sounds and syllables are timed – offer a complementary and potentially more robust dimension for identifying speakers. The authors draw an analogy to gait recognition, where individual movement patterns are highly distinctive. They aim to explore whether idiosyncratic timing in speech, even within a linguistically homogeneous group, can be acoustically measured and used for speaker identification.
Methods
The core of the methodology involves analyzing speech rhythm using various temporal measurement techniques. The authors created the TEVOID-Corpus, featuring spontaneous and read speech from speakers of Zurich German.
Key techniques include:
- Segment Annotation: Manual annotation of speech segments (phonemes) using SAMPA notation, followed by automatic derivation of consonantal and vocalic intervals.
- Rhythmic Measures: A suite of measures are employed, including:
- Ratio Measures: * %V: Percentage of speech that is vocalic.
- %VO: Percentage of speech that is voiced.
- Variability Measures: * VarcoV: Rate-normalized standard deviation of vocalic intervals.
- nPVI-V: Average difference between consecutive vocalic intervals.
- VarcoC: Rate-normalized standard deviation of consonantal intervals.
- rPVI-c: Average difference between consecutive consonantal intervals.
- VarcoPeak: Coefficient of variation of peak-to-peak interval durations (derived from amplitude envelope, often representing syllabic peaks).
- Statistical Analysis: ANOVA is used to determine significant differences in these measures across speakers.
Datasets
- TEVOID-Corpus: A newly introduced corpus specifically designed for this research.
- Content: 8 speakers (initially) producing spontaneous speech (16 sentences each, 10-45 syllables per sentence) from interview situations. Read speech versions of these sentences are also planned.
- Speakers: Homogeneous group of Zurich German native speakers, aged 20-30, with no significant sociolinguistic variability. Gender is the only independent variable considered initially.
- Recording: Sound-treated booth, 44.1 kHz sampling rate, 16-bit.
Key Insights and Ideas
- Time as a Speaker Fingerprint: The paper strongly advocates for temporal speech characteristics as a significant, yet under-explored, dimension for speaker individuality, analogous to physical traits influencing spectral features.
- Rhythm Reveals Idiosyncrasy: Even within a linguistically uniform group (Zurich German), significant rhythmic differences exist between speakers, suggesting idiolectal timing patterns.
- Vocalic and Peak Timing are Key: The study highlights that the timing of vocalic segments (%V, %VO, nPVI-V) and the intervals between syllabic peaks (VarcoPeak) are particularly sensitive to speaker differences.
- Consonantal Timing Less Distinctive (Initially): Measures related to consonantal intervals (VarcoC, rPVI-c) showed less significant variability between speakers in this initial study.
- %V and %VO are Not Highly Correlated: The finding that speakers with high vocalic proportions don't necessarily have high voiced proportions suggests distinct control over these two aspects of speech timing.
Key Results
- Significant Speaker Differences: ANOVA results showed highly significant differences between speakers for %V (F[7, 120]=3.86, p<0.001) and %VO (F[7, 120]=9.42, p<0.001).
- Vocalic Variability: nPVI-V showed a significant effect for between-speaker variability (F[7,120]=2.37; p=0.026), and VarcoV was marginally significant (F[7,120]=2.1; p=0.048).
- Peak-to-Peak Interval Variability: VarcoPeak revealed a highly significant effect for between-speaker variability (F[7, 120]=3.1, p=0.004).
- Consonantal Variability: No significant effects were found for VarcoC or rPVI-c.
- Weak Correlation: The relationship between %V and %VO was poor but highly significant (R²=0.16; F[1,126]=24.8; p<0.001), indicating they capture different aspects of timing.
Limitations and Future Work
- Limited Speaker Sample: The initial study is based on only 8 speakers, and the corpus is still growing.
- Spontaneous Speech Focus: While spontaneous speech is rich in idiosyncratic information, it lacks direct comparability in terms of content. Future work will involve analyzing read speech to control for segmental content and phonotactic sequencing.
- Consonantal Detail: The authors acknowledge that consonantal intervals might be more informative if differentiated by consonant type (e.g., plosives vs. nasals) rather than treated as a single category.
- Within-Speaker Variability: Future work will investigate sources of within-speaker variability, such as voice disguise, and between-speaker similarity, like imitation.
- Anatomical vs. Learned Behavior: The paper discusses whether observed temporal differences stem from anatomical constraints or learned articulatory habits, a distinction that future research aims to clarify.
Summary updated