Using Morphological Data in Language Modeling for Serbian Large Vocabulary Speech Recognition.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Speech Recognition

Background:

• Highly inflective languages, such as Serbian, present challenges for automatic speech recognition (ASR) due to extensive use of suffixes.
• Morphological richness often leads to recognition errors, even when the correct word lemma is identified, due to incorrect word endings.
• Contextual limitations in language model training corpora exacerbate these errors in large vocabulary systems.

Purpose of the Study:

• To investigate the impact of incorporating morphological word categories into language models for ASR.
• To evaluate the effectiveness of this approach in reducing word error rates (WER) and perplexity.
• To demonstrate the benefits for both n-gram and neural network-based language models.

Main Methods:

• Assigning morphological categories (word type, case, number, gender) to words in the system vocabulary.
• Developing and evaluating language models that utilize these additional word features.
• Comparing the performance of the proposed system against a baseline system using standard metrics like WER and perplexity.

Main Results:

• Significant improvements in ASR performance were observed when incorporating morphological features.
• Reductions in word error rates and perplexity were achieved for both n-gram and neural network language models.
• The approach effectively addressed common errors related to incorrect word endings in morphologically rich languages.

Conclusions:

• Integrating morphological information into language models is a viable strategy for enhancing ASR accuracy in languages with rich morphology.
• This method offers a practical solution for reducing tedious errors in large-vocabulary ASR systems, particularly for dictation.
• The proposed approach is applicable to Serbian and other languages exhibiting similar morphological characteristics.