Scholarlink

<p>Recent research shows significant progress in Automatic Speech Recognition (ASR) using advanced transformer models across various languages. However, these models struggle with regional languages like Bengali (Bangla), Hindi, and Japanese, which have many unique characters. Traditional ASR systems perform well on clean, well-annotated datasets, but real-world audio often contains background noise, challenging model accuracy. To address this, the proposed study incorporates a noise distillation block into the transformer architecture to enhance performance in noisy environments. The experiment added background noise to audio samples and applied the noise distillation block during testing to suppress it. Using the Bengali openSLR corpus, the model achieved a Word Error Rate (WER) of 0.2% and Character Error Rate (CER) of 0.08% on clean data. Under noisy conditions, WER and CER rose to 31.84% and 18.28%, respectively. With the noise distillation block, they improved to 17.32% and 9.16%. To further validate results, the English LibriSpeech dataset was also used. These results confirm the block’s effectiveness in noisy ASR scenarios.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Noise-augmented transformer-based automatic speech recognizer using a novel noise distillation system

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Bachchu Paul,
  • Santanu Phadikar,
  • Utpal Nandi

摘要

Recent research shows significant progress in Automatic Speech Recognition (ASR) using advanced transformer models across various languages. However, these models struggle with regional languages like Bengali (Bangla), Hindi, and Japanese, which have many unique characters. Traditional ASR systems perform well on clean, well-annotated datasets, but real-world audio often contains background noise, challenging model accuracy. To address this, the proposed study incorporates a noise distillation block into the transformer architecture to enhance performance in noisy environments. The experiment added background noise to audio samples and applied the noise distillation block during testing to suppress it. Using the Bengali openSLR corpus, the model achieved a Word Error Rate (WER) of 0.2% and Character Error Rate (CER) of 0.08% on clean data. Under noisy conditions, WER and CER rose to 31.84% and 18.28%, respectively. With the noise distillation block, they improved to 17.32% and 9.16%. To further validate results, the English LibriSpeech dataset was also used. These results confirm the block’s effectiveness in noisy ASR scenarios.