Optimizing ASR Models through Contrastive Learning-Based Audio Filtering
Optimizing ASR Models through Contrastive Learning-Based Audio Filtering
Authors:
Prof. Shivani.V. Khedkar1, Mr. Rushikesh. G. Kothare2, Mr. Rohit. S. Ambhore3, Mr. Nitin. R. Diware4, Mr.Pranay. S. Awachar5,
1 Professor, Department of Computer Science Engg,
2 Students, Department of Computer Science Engg, Dr. V. B. Kolte College of Engineering, Malkapur, India
3 Students, Department of Computer Science Engg, Dr. V. B. Kolte College of Engineering, Malkapur, India
4 Students, Department of Computer Science Engg, Dr. V. B. Kolte College of Engineering, Malkapur, India
5Students, Department of Computer Science Engg, Dr. V. B. Kolte College of Engineering, Malkapur, India
Abstract:
This paper presents a novel methodology for enhancing Automatic Speech Recognition (ASR) performance by utilizing contrastive learning to filter synthetic audio data. We address the challenge of incorporating synthetic data into ASR training, especially in scenarios with limited real- world data or unique linguistic characteristics. The method utilizes a contrastive learning model to align representations of synthetic audio and its corresponding text transcripts, enabling the identification and removal of low-quality samples that do not align well semantically. We evaluate the methodology on a medium-resource language across two distinct datasets: a general-domain dataset and a regionally specific dataset characterized by unique pronunciation patterns. Experimental results reveal that the optimal filtering strategy depends on both model capacity and dataset characteristics. Larger models, like Whisper Large V3, particularly benefit from aggressive filtering, while smaller models may not require such stringent filtering, especially on non- normalized text. This work highlights the importance of adjusting synthetic data augmentation and filtering to specific model architectures and target domains. The proposed method, robust and adaptable, enhances ASR performance across diverse language settings. We have open-sourced the entire work, which includes 140 hours of synthetically generated Portuguese speech, as well as the pipeline and parameter settings used to create these samples. Additionally, we provide the fine-tuned Whisper models and the code required to reproduce this research. ai/semantic_audio_filtering.