Abstract
This paper describes a digital segmentation algorithm designed to partition connected natural speech into linguistically relevant units. By analyzing time-domain parameters pitch-synchronously, the algorithm identifies stationary segments, where the vocal tract maintains a target position, and dynamic segments, representing transitions between these targets. This approach addresses the inherent redundancy of acoustic signals and provides a structured foundation for phoneme recognition. Experimental results on a corpus of 200 German words show a boundary detection error rate of only 1.3%.
Introduction
The primary difficulty in automatic speech recognition (ASR) lies in the transition from continuous acoustic signals to discrete symbolic representations like phonemes. Raw acoustic data is highly redundant and lacks clear, inherent boundaries. Traditional methods often relied on simple energy thresholds or voiced/unvoiced detection, which frequently failed to capture the nuances of connected speech. Wolfgang J. Hess proposes a segmentation strategy based on the physical behavior of the human vocal tract, which moves between discrete articulatory targets via continuous transitions.
Methods
The proposed system follows a multi-stage processing pipeline to extract features and define segments.
The algorithm operates in three main stages:
- Pre-segmentation: The signal is first classified into pause vs. speech and voiced vs. voiceless segments using noise-level histograms. This "rough cut" simplifies the search space for more complex parameters.
- Parameter Extraction: Three primary time-domain parameters are measured pitch-synchronously, meaning they are aligned with the fundamental frequency of the voice. These include $L_S$, the absolute average level of the signal; $L_D$, the level of the differenced signal which highlights high-frequency energy; and $Q_S$, the ratio of $L_D$ to $L_S$. $Q_S$ is a clever proxy for spectral behavior, allowing the system to distinguish between vowels, nasals, and fricatives without the computational cost of a full Fourier transform.
- Segmentation Logic: The most novel contribution is the Segment Length Function (SLF). Instead of looking for absolute changes, the algorithm calculates how long a segment can remain stable before a parameter shifts by a "major" relative amount. The signal is broken into tiny "microsegments" (2.5 to 5 ms), and the SLF determines the boundaries of "minimal segments" by looking both forward and backward in time. These segments are then labeled as Stationary (STA), representing a held vocal tract position, or Dynamic (DYN), representing a transition.
Results
The algorithm was tested on a dataset of 200 isolated German words spoken by seven different speakers. The performance was benchmarked against manual expert segmentation. The system achieved a high degree of precision:
- Missed Boundaries: 1.3%
- Missed Stationary Segments: 0.9% (mostly very short glides)
- Insertion Errors (Over-segmentation): 2.1% (usually caused by splitting a single phoneme into two stationary parts)
- Voiced/Unvoiced Accuracy: Errors contributed only 0.3% to the total score, showing the pre-segmentation was highly reliable.
The results indicate that the "Stationary vs. Dynamic" distinction successfully captures the underlying articulatory intent, even when acoustic boundaries are blurred in connected speech.
Conclusion
By prioritizing the rate of change in the vocal tract over absolute energy levels, Hess's algorithm provides a robust framework for speech segmentation. While challenges remain—particularly with the detection of plosives and handling extreme speaker variability such as slurring—the time-domain approach significantly reduces the data redundancy problem for subsequent recognition stages. Future work aims to extend these results from isolated words to rapid, conversational connected speech.
Summary updated