Abstract
Traditional speech recognition systems often rely on fixed-length time windows for analysis, which can lead to parameter instability due to the stochastic nature of speech. This paper introduces the concept of "microphonemes"—dynamic segments synchronized with individual pitch periods. By aligning segmentation with the physical properties of the vocal cords and the biological response of the human ear, this approach provides a more stable foundation for speech parameterization and recognition.
Introduction: The Problem with Fixed Windowing
In the mid-1970s, the prevailing standard for speech recognition systems involved applying Fast Fourier Transforms (FFT) to fixed time intervals, typically ranging from 10 to 20ms. However, speech is not a perfectly periodic signal; it is a complex process with varying amplitudes and waveforms. Using arbitrary fixed windows often results in a "dispersion of parameters," where the resulting data becomes noisy and inconsistent. This inconsistency makes it difficult to reliably identify phonemes or distinguish between different speakers.
The challenge lies in finding a "primary segment" that is mathematically stable, computationally extractable, and biologically relevant. The paper argues that segmentation should be driven by the physics of the human larynx rather than the arbitrary clock cycles of a computer.
Methodology: Defining and Extracting Microphonemes
The author proposes the microphoneme as the fundamental unit of segmentation. A microphoneme is defined by the duration of a single pitch period—the time between successive vibrations of the vocal cords—typically lasting between 3 and 10ms.
System Architecture
The proposed methodology was integrated into a broader processing chain designed to handle quasi-periodic signals.
The extraction process, implemented in Fortran for the H-6030 computer, follows a multi-stage algorithm:
- Initial Classification: The signal is sampled at 12 kHz and categorized into silence, noise, or quasi-periodic (voiced) segments based on amplitude and zero-crossing thresholds.
- Extreme Detection: Within voiced segments, the algorithm identifies "significant extremes"—local maxima and minima representing the core of the vocal cord's vibration.
- Boundary Refinement: Using a "unilaterally convergent" correction process, the algorithm pinpoints the exact start and end of each microphoneme. It specifically looks for the point of minimum envelope or a zero-crossing immediately preceding the maximum turbulence.
- Bio-mimicry: The logic is grounded in an analog model of the human ear, specifically how the cochlea septum responds to auditory pressure.
Visualizing Segmentation
By syncing segments to pitch periods, the resulting parameters become "univocal" or unambiguous, providing a clearer representation of the speech signal.
Key Insights and Results
The study evaluated the algorithm using a specialized dataset of 10 Polish words (e.g., ala, rojal, rzeka, wych, jodla, nora, ryba, masz, foka). Key findings include:
- Parameter Stability: Microphonemes satisfied the mathematical conditions for stable speech parameterization significantly better than traditional fixed-length segments.
- Automated Precision: Interestingly, the automated algorithm often identified more microphonemes than human experts could manually detect on an oscilloscope, capturing nuances invisible to the naked eye.
- Computational Feasibility: Despite the complexity of multi-stage refinement, the process was successfully automated, proving that pitch-synchronous analysis is viable for systematic speech processing.
Conclusion
The shift from fixed-window analysis to pitch-synchronous microphonemes represents a move toward more biologically and physically accurate speech modeling. While hardware constraints of the era limited testing to shorter sequences, the results suggest that microphonemes provide a superior basis for information recognition, speaker identification, and noise reduction. Future work should focus on applying this method to larger corpora to further validate its effectiveness for real-time automatic speech recognition.
Summary updated