TL;DR
This paper compares the performance of two speech analysis software packages, PRAAT and CSL, in analyzing pathological voice data. The authors found that while both tools can extract perturbation parameters like jitter, PRAAT offers advantages in automation and potentially more accurate pitch detection, especially for complex voice signals.
What Problem Does This Paper Address?
Accurate analysis of pathological voice signals is crucial for diagnosing laryngeal disorders. This analysis often relies on extracting acoustic parameters, particularly those related to voice perturbation (variations in pitch and amplitude). While the Computerized Speech Lab (CSL) has been a long-standing standard in this field, it's a commercial, integrated system. PRAAT, a free and open-source alternative, is continuously being developed. This paper aims to compare the perturbation analysis capabilities of these two tools, specifically focusing on pitch detection and its impact on jitter measurements, to understand if PRAAT can serve as a viable alternative or complement to CSL.
Methods
The core of this study involves comparing the output of PRAAT and CSL when analyzing the same set of voice samples. The authors focused on perturbation analysis, specifically examining parameters related to fundamental frequency (F0) variations, such as jitter (cycle-to-cycle variation in frequency) and the difference between maximum and minimum F0.
PRAAT's pitch detection algorithm, which uses autocorrelation and a "sinx/x" interpolation technique for improved resolution, was employed. The study also highlights PRAAT's scripting capabilities, allowing for automated analysis of multiple files. CSL's analysis was used as a benchmark. The authors generated synthetic sine waves with controlled noise and harmonic components to test the pitch detection accuracy of both systems under varying conditions. They also analyzed a subset of pathological and normal voice samples from the "Voice disorders database."
Datasets
The study utilized two types of datasets:
- Synthetic Datasets: Pure 113Hz and 226Hz sine waves were generated with varying conditions, including the addition of Gaussian noise and a harmonic component (266Hz to 133Hz at a 2:1 ratio). These were used to validate the pitch detection accuracy of PRAAT.
- "Voice Disorders Database": This database, originally from the Massachusetts Eye and Ear Infirmary, contains both normal and pathological voice samples. The authors converted the CSL-formatted (.nsp) files to standard wave format for analysis with PRAAT. They selected 25 pathological and 12 normal voice samples for their comparative analysis.
Key Insights and Ideas
- PRAAT's Automation Advantage: PRAAT's scripting capabilities are a significant advantage for analyzing large datasets of voice samples, making it more efficient than manual analysis or less automated commercial software.
- Pitch Detection Nuances: The accuracy of pitch detection is paramount for reliable perturbation analysis. While both tools generally agree on mean F0, significant differences can emerge in jitter values, suggesting variations in how they handle unstable or complex pitch periods.
- Impact of Noise and Harmonics: Synthetic tests showed that adding Gaussian noise increased jitter and standard deviation of F0, as expected. PRAAT demonstrated an ability to accurately detect pitch periods even when a harmonic component was present.
- Potential for CSL's Limitations: The authors suggest that CSL might not use an interpolation method for pitch extraction, potentially leading to errors at the end of voice samples. They also note that CSL's "Voice disorders database" results sometimes show very large differences between maximum and minimum F0 that are not perceptually evident, hinting at potential issues in their analysis.
Key Results
- Jitter Differences: Table 3 and Appendix 2 show that while mean F0 values are often similar between CSL and PRAAT, jitter values can differ significantly, with CSL often reporting higher jitter for pathological voices.
- Number of Pitch Periods: PRAAT generally reported a smaller number of pitch periods compared to CSL for the same voice samples.
- Synthetic Wave Validation: For synthetic pure sine waves, both CSL and PRAAT provided comparable results for jitter and standard deviation of F0. The addition of noise increased these values, but the differences between CSL and PRAAT remained small.
- Harmonic Component Handling: PRAAT was able to perfectly find exact pitch periods even when a harmonic component (266Hz to 133Hz) was added to the signal.
Limitations and Future Work
- Shimmer Parameter Exclusion: The study focused solely on frequency perturbation (jitter) and excluded shimmer (amplitude perturbation) because PRAAT does not have a direct function to obtain shimmer parameters.
- CSL Algorithm Uncertainty: The authors acknowledge they do not fully know CSL's algorithms, which limits a definitive explanation for the observed differences.
- Need for More Validation: The authors explicitly state that "more validation experiments can be needed in order to prove the shortcoming in CSL."
- Future Directions: The paper suggests that PRAAT can be used for other analyses, such as time-frequency analysis, which could be beneficial for laryngeal diagnosis. Further research could explore shimmer analysis with PRAAT and more extensive validation against CSL and other established methods.
Summary updated