Real and synthetic Punjabi speech datasets for automatic speech recognition

Area of Science:

• Computational Linguistics
• Speech Processing

Background:

• Automatic speech recognition (ASR) systems require large annotated datasets for robust performance.
• High-resource languages like English dominate available ASR datasets, creating a resource gap for low-resource languages.
• Punjabi, despite its numerous speakers, lacks sufficient high-quality annotated speech datasets.

Purpose of the Study:

• To address the scarcity of labeled Punjabi speech data.
• To facilitate the development of accurate Punjabi automatic speech recognition systems.
• To bridge the resource gap for low-resource language ASR.

Main Methods:

• Introduction of three novel Punjabi speech datasets: Punjabi Speech (real recordings), Google-synth (Google TTS), and CMU-synth (Clustergen model).
• Punjabi Speech dataset includes read speech recorded in diverse environments (studio and open settings).
• Synthesized datasets created using Google's Punjabi text-to-speech and CMU's Clustergen model within the Festival system.

Main Results:

• The creation of three distinct, labeled Punjabi speech datasets.
• Availability of both real and synthesized speech data for ASR model training.
• Establishment of resources to support research in Punjabi speech recognition.

Conclusions:

• The newly introduced datasets provide crucial resources for advancing Punjabi ASR.
• These datasets will enable the training of more accurate and robust Punjabi speech recognition models.
• This initiative helps mitigate the data scarcity issue for Punjabi in the field of ASR.