Microphonemes As Fundamental Segments Of Speech Wave Primary Segmentation - Automatic Searching For Microphonemes

TL;DR

This paper proposes a novel approach to speech signal segmentation by introducing the concept of "microphonemes" as fundamental units. This method aims to improve speech recognition, compression, and speaker identification by providing a more optimal and dynamic segmentation strategy than traditional fixed-interval approaches.

What Problem Does This Paper Address?

The paper tackles the challenge of effectively segmenting speech waves for various applications like recognition, compression, and speaker identification. Existing methods often rely on fixed time intervals (e.g., 10-20 ms) for spectral analysis, which can lead to inaccuracies due to the inherent variability of speech parameters within these segments. This inflexibility hinders the development of robust and efficient automatic speech processing systems. The authors argue that a more dynamic and optimal segmentation approach, based on fundamental units of speech, is crucial for overcoming these limitations.

Methods

The core of the proposed method is the introduction of "microphonemes" as the fundamental segments of voiced speech. These microphonemes are derived from the concept of quasi-periodicity within speech segments and are inspired by an analog model of the human ear, particularly the function of the stirrup bone and cochlea.

The segmentation algorithm operates in several stages:

  • Primary Segmentation: This involves identifying voiced segments of the speech wave using amplitude and frequency criteria. The algorithm distinguishes between silence, noise, and quasi-periodic signals.
  • Microphoneme Search: Within voiced segments, the algorithm searches for "significant extremes" (maxima and minima) that mark the boundaries of microphonemes. This process is inspired by Reddy's Pitch Period Determination algorithm but includes modifications for Polish speech and the concept of microphonemes.
  • Correction: A crucial step involves correcting erroneously calculated microphoneme ends. This is achieved by comparing the indicated ends with expected ends and applying a convergence process to refine the boundaries.

The authors emphasize that microphonemes are defined by their pitch period, which is related to the larynx tone (TK), and are considered fundamental units for analysis.

Datasets

Not discussed in the paper. The authors mention using the ZAM-41 computer for initial research and later processing on an H-6030 computer with a Fortran program. They also mention analyzing "10 single words" for initial testing.

Key Insights and Ideas

  • Microphonemes as Optimal Segments: The paper posits that microphonemes, derived from the quasi-periodic nature of speech and inspired by the human ear's spectral analysis, offer a more optimal segmentation unit than fixed-time intervals.
  • Dynamic Segmentation: The proposed method moves away from rigid time-based segmentation towards a dynamic approach that adapts to the inherent variations in speech.
  • Bridging Phonemes and Speech Analysis: The authors suggest that microphonemes can serve as a bridge between phoneme-level analysis and broader speech wave analysis, potentially simplifying recognition and compression tasks.
  • Human Ear Analogy: The inspiration drawn from the human ear's analog model provides a biologically plausible basis for the proposed segmentation strategy.

Key Results

The paper describes the algorithm and its conceptual underpinnings rather than presenting extensive quantitative results. However, they state that their initial analysis of "10 single words" yielded satisfactory results, finding more microphonemes than a human observer would typically identify, especially at word endings. They also note that the algorithm's efficiency was limited by the ZAM-41 computer's slowness, prompting a move to a more powerful system for further testing.

Limitations and Future Work

  • Computer Limitations: The primary limitation acknowledged is the slowness of the ZAM-41 computer, which restricted the length and complexity of the signals that could be analyzed. This necessitated the development of a new Fortran program for the H-6030 computer to enable more extensive testing.
  • Signal Length: More exacting research on longer signals is needed to fully validate the algorithm's performance.
  • Comparison with Existing Methods: The authors express a desire to compare their algorithm's effectiveness against previously used frequency analysis methods for speech sounds once the new program is operational.
  • Generalizability: While tested on Polish speech, further validation across different languages and speaker populations would be beneficial.

Summary updated

BibTeX
@inproceedings{1975microphonemes,
  title = {{Microphonemes As Fundamental Segments Of Speech Wave Primary Segmentation - Automatic Searching For Microphonemes}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1975},
}
Figure