Heterophonic speech recognition using composite phones.

Area of Science:

• Computational Linguistics
• Speech Recognition Technology
• Natural Language Processing

Background:

• Heterophones, words with identical spelling but varied pronunciation, present significant challenges in training Automatic Speech Recognition (ASR) systems.
• Current ASR methods struggle with the inherent ambiguity of heterophonic languages, particularly Arabic, due to orthographic representations lacking phonetic detail.

Purpose of the Study:

• To introduce and evaluate a novel pronunciation unit, the Composite Phoneme (CP), designed to address the challenges posed by heterophonic languages in ASR.
• To develop algorithms for generating CP pronunciations from Modern Orthography (MO) and compare their performance against traditional methods.

Main Methods:

• Developed Composite Phonemes (CP) as a set of alternative phoneme sequences, focusing on consonant-centric units that incorporate short vowels and gemination absent in Modern Orthography (MO).
• Created algorithms to generate CP pronunciations from MO and Simple Phoneme (SP) pronunciations from Classical Orthography (CO).
• Investigated and compared the performance of ASR systems using CP, SP, Undiacritized Grapheme (UG), and Diacritized Grapheme (DG) on the A-SpeechDB corpus.

Main Results:

• ASR systems utilizing CP and SP demonstrated superior performance over UG and DG approaches.
• For an 8,000-word MO vocabulary, CP achieved a Word Error Rate (WER) of 11.78%, outperforming SP_M (12.64%) and SP_A (13.59%).
• For a 24,000-word MO vocabulary, CP's WER was 13.69%, compared to SP_M's 15.08% and SP_A's 16.86%.

Conclusions:

• Composite Phonemes (CP) offer a more effective approach for ASR in heterophonic languages compared to grapheme-based methods.
• While SP shows better performance with uniform statistical models, CP excels in context-independent phone scenarios.
• The CP approach alleviates the necessity for diacritization, simplifying the ASR training process for Arabic.