A Study of Speech Recognition for Kazakh Based on Unsupervised Pre-Training

Area of Science:

• Speech Recognition
• Natural Language Processing
• Machine Learning

Background:

• Low-resource languages like Kazakh face challenges in developing effective speech recognition systems due to limited paired data.
• Unsupervised pre-training methods have shown promise for low-resource ASR but are underutilized for Central and West Asian languages.

Purpose of the Study:

• To improve the performance of automatic speech recognition (ASR) for low-resource languages.
• To adapt and enhance the wav2vec2.0 model for better speech representation learning.
• To investigate the effectiveness of multi-language pre-training and speech synthesis for ASR data augmentation.

Main Methods:

• An improved wav2vec2.0 model, termed wav2vec-F, was developed by integrating a Factorized TDNN layer.
• Unsupervised pre-training was employed on large unlabeled audio datasets to learn speech representations.
• A cross-language ASR task was optimized using noise contrastive binary classification, incorporating speech synthesis for data enhancement.

Main Results:

• wav2vec-F effectively leverages unlabeled data from non-target languages, with multi-language pre-training outperforming single-language pre-training.
• Speech synthesis as a data enhancement method significantly improved ASR performance.
• A word error rate reduction of 1.9% on Librispeech's test-clean dataset was observed.
• On the Kazakh KSC test set, Kazakh-only pre-training reduced the word error rate by 3.8%.

Conclusions:

• The proposed wav2c-F model and multi-language pre-training strategy are effective for low-resource ASR.
• Speech synthesis is a highly beneficial data augmentation technique for improving ASR accuracy.
• The approach achieved comparable results to previous end-to-end models with significantly less labeled data, demonstrating its efficiency for low-resource scenarios.