TL;DR
This paper introduces a novel semi-automatic speech analysis environment designed to robustly measure acoustic voice parameters in continuous speech. By effectively segmenting speech into voiced, unvoiced, and silence components, the system significantly improves the reliability of voice analysis, making it more practical for studying vocal loading in real-world scenarios.
What Problem Does This Paper Address?
Understanding vocal loading – the strain placed on the voice from extensive use – is crucial for occupational health, especially in voice-intensive professions. While previous research has explored this, analyzing continuous speech presents significant technical challenges. The dynamic nature of natural speech, with rapid variations and short phonetic segments, makes it difficult to extract reliable acoustic measures like fundamental frequency (F0) and sound pressure level (SPL). Existing methods often rely on sustained vowels or isolated words, which may not accurately reflect real-world vocal behavior. This paper tackles the problem of creating a robust and efficient system for analyzing continuous speech, minimizing user intervention and handling large datasets, thereby enabling more accurate and practical studies of vocal loading.
Methods
The core of the system lies in its semi-automatic speech analysis pipeline, which involves three main stages:
- Voice Activity Measures and Classification: The system first classifies speech signals into segments of silence, voiced speech, and unvoiced speech. This is achieved by analyzing speech in 20ms non-overlapping frames, extracting three key parameters:
- Sound Pressure Level (SPL): Measures the loudness of the speech.
- Fundamental Frequency (F0): The rate of vocal fold vibration, indicating pitch.
- Stationarity Measure (using Akaike Information Criterion): This novel approach quantifies the predictability of the speech signal. It leverages the fact that voiced speech is generally more stationary than unvoiced speech or noise. The formula provided (c = NrTr / r²(0)) quantifies this stationarity.
- Post-Classification Refinement: To address potential classification errors, a post-classification procedure is applied. This involves thresholding the lengths of speech and silence segments. Segments shorter than a defined threshold are merged into the adjacent segment type, ensuring more consistent and accurate segment identification.
- Voice Quality Measures: Once segments are classified, various voice quality measures are extracted. In addition to F0 and SPL, the alpha-ratio is calculated. This is defined as the ratio of energy in the signal above 1 kHz to the energy below 1 kHz, providing insights into spectral characteristics.
The system is implemented using LabVIEW for the user interface and MATLAB for mathematical programming, running on standard MS-Windows computers.
Datasets
The speech material used for evaluation was derived from customer service recordings at a call center. This included recordings from 36 subjects (27 female, 9 male). For the specific evaluation presented in the paper, a subset of these voices was used. The recordings were made on DAT tapes with a sampling frequency of 44.1 kHz.
Key Insights and Ideas
- Automatic Pre-segmentation is Key: The paper strongly suggests that accurately segmenting continuous speech into voiced, unvoiced, and silence is a critical prerequisite for robust acoustic voice analysis. This pre-segmentation significantly improves the reliability of subsequent measurements.
- Stationarity as a Discriminator: The use of the Akaike Information Criterion to derive a stationarity measure is a novel approach for distinguishing between voiced and unvoiced speech, offering a more sophisticated method than simple energy thresholds.
- User Variance is Manageable: The study demonstrates that while some user-induced variance exists (intra-operator variance), it is generally small (around 1% of the mean for most parameters). This suggests the system is relatively intuitive to use.
- Robustness to Noise: The system shows good performance even in the presence of background noise, with voiced/unvoiced classification remaining effective down to low Signal-to-Noise Ratios (SNR). This is crucial for real-world applications where perfect recording conditions are rare.
- Alpha-Ratio Behavior in Noise: The alpha-ratio was observed to increase with decreasing SNR, which is an expected behavior given how noise affects high-frequency components. This highlights that while the system is robust, understanding the behavior of individual measures under different conditions is important.
Key Results
- Intra-operator variance: Standard deviations for most parameters were approximately 1% of the mean or below, indicating high consistency when a single operator uses the system. The 95% confidence interval for measurements was estimated to be around 2% of the measured value.
- Inter-operator variance: While present, inter-operator variance was generally smaller than intra-operator variance for most measures, suggesting that learning the software is straightforward. Some measures, like SPL and segment lengths, were more sensitive to operator choices.
- Robustness to SNR: The system successfully classified voiced/unvoiced segments even with SNR levels as low as 2 dB. The alpha-ratio and SNR estimates were found to be relatively stable across various SNR levels, though the alpha-ratio did increase with decreasing SNR.
- No significant gender differences: In most tests, no significant differences in measurement variance were found between male and female voices, even for extreme F0 values.
Limitations and Future Work
- Minor User-Induced Variance: While generally low, some parameters (e.g., SPL, unvoiced segment length) were more sensitive to operator choices. The paper suggests that the learning process for the software is straightforward, but individual voice characteristics might influence inter-operator variance more than software learning itself.
- Total Recording Time Inaccuracy: The user interface software occasionally exceeded the preset recording time by a few seconds due to synchronization issues with data flux. This was deemed a minor problem.
- Future Work: The authors suggest expanding the software to include automatic inverse filtering of the voice source. They also note that the developed classification algorithm is not limited to vocal loading studies and can be applied to any task requiring voice/unvoiced classification of continuous speech.
Summary updated