Speech sample salience analysis for speech cycle detection

C. Mertens, Francis Grenez, Jean SchoentgenInterspeech 2009

TL;DR

This paper introduces a novel "salience analysis" method for detecting speech cycles, crucial for analyzing vocal perturbations like jitter and tremor. Unlike traditional methods that assume local periodicity, this approach identifies speech cycles by measuring how long a sample remains a maximum within its local context, offering a more robust analysis for pathological speech.

What Problem Does This Paper Address?

Analyzing speech requires understanding its fundamental building blocks, particularly the "speech cycles" within voiced sounds. These cycles are essential for measuring subtle variations in vocal frequency (jitter) and amplitude (shimmer), which can indicate vocal health issues or be used in speech synthesis. Traditional methods for detecting these cycles often rely on identifying peaks in the speech signal and assume these peaks occur at roughly regular intervals. This assumption, however, breaks down for pathological voices where the regularity is disrupted, leading to errors in cycle detection. The paper addresses this by proposing a method that doesn't require prior knowledge of local periodicity or average cycle length, making it more suitable for analyzing a wider range of speech, including disordered voices.

Methods

The core of the paper is the "salience analysis" method. For each sample in a speech signal, salience quantifies the length of the longest interval where that sample is the maximum. This is calculated using a basic algorithm that iteratively considers increasing interval lengths and assigns the maximum length to the sample's salience.

To address boundary effects (where samples near the start or end of the signal might have artificially low salience), two main strategies are explored:

  1. Within-array rotation: The signal array is rotated multiple times so each sample gets a chance to be at the boundary, and the final salience is an average. This is computationally intensive.
  2. Partial salience allocation & Sliding analysis window: A more efficient approach is presented. It involves a sliding window that moves across the signal. Within each window, a "partial salience" is calculated. The final "running salience" for each sample is the maximum of its local saliences across all window positions. This ensures that salience values only increase, reflecting the most prominent local maximum over time.

For actual speech cycle detection, the method is applied to the speech signal after preprocessing (decimation and band-pass filtering). Prominent speech cycle peaks are identified based on high salience values. The selection of relevant peaks is further refined using the "coefficient of variation" of inter-peak durations, aiming to find the sequence of peaks with the most consistent spacing.

Datasets

The paper uses a corpus of seven sustained vowels [a] synthesized by a simulator of disordered voices. These synthetic signals were generated with increasing amounts of jitter to evaluate the method's performance under controlled perturbation conditions.

Key Insights and Ideas

  • Salience as a robust peak descriptor: The concept of salience provides a measure of a sample's prominence that is independent of strict periodicity assumptions, making it more resilient to signal irregularities.
  • Overcoming boundary effects: The sliding window approach effectively mitigates issues caused by signal boundaries, ensuring more reliable salience values for all samples.
  • Salience for cycle detection: High salience values correlate well with important speech cycle peaks, providing a principled way to identify them.
  • Jitter analysis comparison: The salience-based method shows comparable trends to established methods (like Praat) for jitter analysis, suggesting its validity, while offering advantages in robustness.

Key Results

The paper presents a comparison of jitter analysis results between their salience-based method and the Praat software. Table 1 shows that while the mean fundamental frequency and standard deviation of cycle durations are similar, the jitter percentages differ. The authors attribute these differences to Praat's peak detection not always identifying the most salient peaks. The salience analysis method, by not assuming regularity, is presented as an advantage for analyzing pathological voices.

Limitations and Future Work

  • Parameter tuning: The length of the sliding analysis window (N) is a parameter that needs to be chosen based on the expected vocal frequency. While the paper provides a formula (N = 1.5 * F_max / F_min), this still requires some prior knowledge.
  • Computational efficiency: While the sliding window approach is faster than rotation, further speed-up is mentioned as possible by incorporating additional tests on saliences from previous window positions, though these are not detailed.
  • Qualitative differences in jitter: The paper notes that jitter values are not identical between methods, implying that the precise definition and measurement of jitter can vary, and further investigation into these differences might be warranted.
  • Future work: The authors suggest that further optimization of the algorithm for speed is possible. They also imply that the robustness of salience analysis could be beneficial for other signal processing tasks beyond speech cycle detection.

Summary updated

BibTeX
@inproceedings{mertens2009speech,
  author = {C. Mertens and Francis Grenez and Jean Schoentgen},
  title = {{Speech sample salience analysis for speech cycle detection}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure