Quantifying Fundamental Frequency Modulation as a Function of Language, Speaking Style and Speaker

Pablo Arantes, Anders ErikssonInterspeech 2019

TL;DR

This paper introduces a novel method to quantify differences in fundamental frequency (f0) modulation patterns across languages, speaking styles, and individual speakers. They found that speaking style significantly impacts f0 modulation, with spontaneous speech showing greater inter-speaker variability, suggesting its potential utility in forensic phonetics.

What Problem Does This Paper Address?

In forensic phonetics, comparing questioned speech samples (e.g., from a bugged phone call) with reference samples (e.g., from a police interview) can be challenging if the speaking styles differ. This mismatch can lead to false negatives because speakers' vocal characteristics, particularly their f0 modulation, can vary significantly across different speaking styles and emotional states. While previous research has explored global f0 measures (like mean and median), this paper aims to quantify finer-grained differences in f0 modulation patterns, which are often overlooked but could be crucial for robust speaker comparison. They specifically address the gap in understanding how language, speaking style, and individual speaker characteristics interact to shape these f0 modulation patterns.

Methods

The core of this paper is a new methodology for quantifying the similarity between two f0 distributions. They employ the Anderson-Darling (A-D) k-sample test, a statistical measure that, unlike the Kolmogorov-Smirnov test, gives more weight to the tails of distributions. This makes it particularly sensitive to differences in f0 modulation, such as variations in spread, range, skewness, and kurtosis, rather than just central tendency.

Before applying the A-D test, raw f0 values (in Hz) are converted to the OMe scale (log2(f_hz / f_med)). This normalization centers each distribution around the speaker's median f0, effectively removing differences in absolute pitch and focusing solely on the shape of the f0 contour. The resulting f0 contours are then transformed into smoothed histograms (kernel density estimates) to represent the probability density of the f0 data. The A-D statistic (A²) is then calculated between pairs of these smoothed histograms. A higher A² value indicates greater dissimilarity between the f0 distributions.

Statistical analysis involves using Kruskal-Wallis tests and paired t-tests (with Holm correction) to assess the significance of effects of language and speaking style on the mean A² values. They also use ANOVA and one-sample t-tests to investigate speaker variability and compare within-speaker vs. between-speaker differences.

Datasets

The study utilizes a subset of a multilingual corpus previously used for research on lexical stress. This corpus includes recordings in Brazilian Portuguese, British English, Estonian, French, German, Italian, and Swedish. For each language, there are 10 speakers (5 male, 5 female), each contributing speech in three distinct styles: spontaneous interview, sentence reading, and word list reading. The authors note that the speech material for sentence and word list readings was derived from transcriptions of the spontaneous interviews, ensuring identical linguistic content across styles.

Key Insights and Ideas

  • f0 Modulation is a Rich Source of Information: The paper strongly suggests that focusing on the shape of f0 distributions, rather than just average pitch, reveals significant differences related to language, speaking style, and speaker.
  • Speaking Style Matters Significantly: The study demonstrates that speaking style has a profound impact on f0 modulation patterns. This is a crucial finding for applications like forensic phonetics, where style mismatches are common.
  • Spontaneous Speech for Speaker Discrimination: Spontaneous speech (specifically interviews) appears to be the most informative style for distinguishing between speakers based on their f0 modulation patterns, showing the highest inter-speaker variability in f0 distribution similarity.
  • Language-Specific Style Implementation: The way speaking styles influence f0 modulation is not uniform across languages. Some languages show greater differences between interview and read speech, while others show more variation between sentence and word reading.
  • Within-Speaker Variability is Lower Than Between-Speaker Variability: A key finding for forensic applications is that an individual speaker's f0 modulation patterns tend to be more similar across different styles than the patterns of different speakers speaking in the same style. This is a desirable characteristic for a forensic parameter.

Key Results

  • Speaking Style Effects: The A-D test revealed significant effects of speaking style on f0 distribution similarity across all languages.
  • For English, French, Italian, Portuguese, and Swedish, spontaneous interviews showed the highest inter-speaker variability (highest mean A²).
  • For Estonian and German, sentence reading showed the highest inter-speaker variability.
  • Speaker Variability: Statistical analysis indicated that the effect of speaking style on f0 distribution shape variation is largely uniform across speakers within each language.
  • Within- vs. Between-Speaker Variability: A comparison of within-speaker (across styles) and between-speaker (same style) variability showed that within-speaker variability is significantly lower. For example, in English, the mean A² for within-speaker comparisons was 175, while for interview style between-speaker comparisons, it was 1047. This difference was statistically significant.

Limitations and Future Work

  • Single Sample per Style: The authors acknowledge that for the within-speaker analysis, they only had one sample per speaker per style. They suggest that future work should investigate within-speaker variability using multiple samples from the same speaker in the same style to further confirm their findings.
  • Statistical Descriptors: While the A² statistic effectively quantifies overall distribution similarity, the authors propose that future research should delve deeper to identify which specific statistical descriptors (e.g., range, asymmetry, kurtosis) correlate most strongly with this metric.
  • Forensic Validation: While the results are promising for forensic phonetics, further studies are needed to confirm the robustness of this method in real-world forensic scenarios.

Summary updated

BibTeX
@inproceedings{arantes2019quantifying,
  author = {Pablo Arantes and Anders Eriksson},
  title = {{Quantifying Fundamental Frequency Modulation as a Function of Language, Speaking Style and Speaker}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure