Rhythm and formant features for automatic alcohol detection

Florian Schiel, Christian Heinrich, Veronika NeumeyerInterspeech 2010

TL;DR

This paper explores the use of speech rhythm and formant features to automatically detect alcohol intoxication. They found that specific rhythm and formant features can reliably distinguish between sober and alcoholized speech, even suggesting that alcoholized speakers may hyper-articulate to compensate for their state.

What Problem Does This Paper Address?

The ability to automatically detect alcohol intoxication from speech has significant implications for forensic science and public safety, such as in-car alcohol detection systems. However, previous research on this topic has yielded mixed results, likely due to variations in empirical data and methodologies. This paper aims to address this by investigating specific acoustic features – rhythm and formants – within a large, well-characterized corpus of alcoholized and sober speech to identify reliable indicators of intoxication. A key challenge is that alcohol's effects can manifest differently across individuals, making speaker-dependent analysis crucial.

Methods

The researchers analyzed two sets of speech features:

  • Rhythm Features: These were derived from the short-time Root Mean Square (RMS) energy of the speech signal. They developed a method to extract sequences of local minima and maxima from the RMS trajectory. From these sequences, they calculated measures related to the timing and regularity of these energy peaks, including mean and standard deviation of time delays between successive peaks, and differences in RMS values between successive peaks.
  • Formant Features: Formants, which represent vocal tract resonances, were extracted using the SNACK package's formant tracker. They focused on the first four formants (F1-F4) and calculated their median frequencies and quarter-quantile distances within vowel segments. They also investigated the "vowel space" by analyzing the Euclidean distances between vowel centroids (representing /a:/, /i:/, /u:/) and the overall centroid.

Statistical analysis involved repeated measures ANOVA (RM-ANOVA) and mixed models (MM) to identify significant differences between sober and alcoholized speech, accounting for factors like gender, speech style, and vowel class. They also employed a speaker-dependent pattern recognition framework, dividing data into development and test sets to build individual models.

Datasets

The primary dataset used is the Alcohol Language Corpus (ALC). This corpus, collected by the Bavarian Archive for Speech Signals (BAS), includes speech from 77 female and 84 male speakers recorded in both sober and alcoholized conditions. The corpus features various speech styles (read speech, list style, semi-spontaneous, spontaneous, situational commands) and content (digit strings, addresses, tongue twisters, etc.). A subset of 20 speakers was recorded a third time in the same situation as the alcoholized recordings but sober, serving as a control group.

Key Insights and Ideas

  • Rhythm as a Strong Indicator: The study found that specific rhythm features derived from RMS energy are highly sensitive to alcohol intoxication, showing significant differences between sober and alcoholized speech across various conditions.
  • Hyper-articulation Hypothesis: Contrary to the expectation that vowels might become more centralized in the F1/F2 space when intoxicated, the study observed the opposite – a decentralization. The authors hypothesize this is due to subjects attempting to compensate for their impaired state by hyper-articulating, especially when aware they are being tested.
  • Speaker-Dependent Analysis is Crucial: The paper emphasizes that alcohol's effects on speech are highly individual. Therefore, effective detection requires models that can adapt to individual speaker characteristics rather than relying solely on population-level averages.
  • Formant Features Show Mixed Results: While some formant features showed significant differences, their reliability was often dependent on gender, vowel class, or speech style, making them less universally applicable than some rhythm features.

Key Results

  • Rhythm: Three rhythm features (measures A, B, C, D, E, F, G, H, I are discussed, with A, B, C, D, E, F, G, H, I showing significant differences) showed a significant increase with alcoholization, independent of gender and speech style (p<0.0001 for most).
  • Formants:
    • F1m (median F1 frequency) significantly increased in alcoholized speech (p<0.0001), particularly for the vowel /a:/.
  • F4qq (quarter-quantile distance of the fourth formant) showed a highly significant increase across genders, vowel classes, and speech styles (p=0.0040).
  • ED (Euclidean distance measure for vowel space centralization) also showed a significant change (p<0.0001), with sober speech appearing more centralized, contradicting the initial hypothesis.
  • Overall: The study identified 3 rhythm and 3 formant features with high potential for detecting intoxication.

Limitations and Future Work

  • Limitations:
    • The ALC corpus does not contain speech data with varying BAC levels from the same speaker, preventing direct correlation tests between feature values and BAC within a single speaker.
  • The observed hyper-articulation in the test setting might not reflect natural intoxication scenarios.
  • Future Work:
    • Investigate whether incorporating the identified rhythm and formant features can improve upon existing methods (like those using MFCCs, which achieved an average prognosis rate of 77% on the same database) for automatic alcohol detection.
  • Further explore the hyper-articulation phenomenon and its implications for detection models.
  • Conduct experiments with varying BAC levels to establish direct correlations.

Summary updated

BibTeX
@inproceedings{schiel2010rhythm,
  author = {Florian Schiel and Christian Heinrich and Veronika Neumeyer},
  title = {{Rhythm and formant features for automatic alcohol detection}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure