A probabilistic framework for landmark detection based on phonetic features for automatic speech recognition

Area of Science:

• Speech Recognition
• Phonetics
• Acoustic Phonetics

Background:

• Traditional speech recognition systems often rely on Mel-frequency cepstral coefficients (MFCCs).
• Acoustic parameters (APs) offer a different approach by directly capturing phonetic features.
• Landmark-based methods focus on identifying key points in speech signals.

Purpose of the Study:

• To present a novel probabilistic framework for landmark-based speech recognition.
• To utilize acoustic parameters (APs) for detecting phonetic landmarks related to manner features.
• To evaluate the performance of this framework against established methods.

Main Methods:

• Developed a probabilistic framework for detecting multiple landmark sequences in continuous speech.
• Employed acoustic parameters (APs) capturing manner-based phonetic features (syllabic, sonorant, continuant).
• Used binary classifiers for probabilistic landmark detection and constrained sequences with pronunciation models.

Main Results:

• The probabilistic landmark detection system utilizing APs demonstrated competitive performance.
• Comparison with MFCC-based probabilistic and Hidden Markov Model (HMM) systems was conducted.
• The framework's ability to exploit feature sufficiency and invariance was highlighted.

Conclusions:

• The proposed probabilistic framework offers a viable alternative for speech recognition.
• Acoustic parameters provide valuable information for phonetic landmark detection.
• Further research can explore integrating this framework with other speech recognition paradigms.