Scholarlink

<p>Voice assistants are an integral part of our daily lives, yet a significant problem persists: false triggering. Voice assistants often trigger when they hear phrases similar to their wake words. Prior research focuses on training CRNN (Convolutional Recurrent Neural Networks) models with Mel Frequency Cepstral Coefficients (MFCCs) or Log Filter Bank Energies (LFBEs) to improve accuracy and reduce either false accepts or false rejects. The existing research addresses only one of the three significant problems—false predictions—among false predictions, model complexity, and latency. It does not provide an optimized solution to resolve all three issues. This paper proposes a novel wake-word detection model using the Turbo-CRNN architecture with Linear Frequency Cepstral Coefficients (LFCCs) to minimise the adverse effects of the above problems. This paper experimented with multi-head attention, regularizers, and optimizers to determine the best possible model for this application to minimize false positives and negatives. This proposed custom CRNN model achieved higher accuracy, lower model complexity (87k parameters), significantly fewer false predictions, and very low latency (around 80 ms) compared to the CNN, DNN, and TinyCRNN baselines.</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

TurboCRNN: a robust and efficient architecture for wake word detection

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Nitin Nataraj,
  • Tejeswar A. Ramachandran,
  • Dhairya Lalwani,
  • Mohanaprasad Kothandaraman,
  • V. Soni Ishwarya

摘要

Voice assistants are an integral part of our daily lives, yet a significant problem persists: false triggering. Voice assistants often trigger when they hear phrases similar to their wake words. Prior research focuses on training CRNN (Convolutional Recurrent Neural Networks) models with Mel Frequency Cepstral Coefficients (MFCCs) or Log Filter Bank Energies (LFBEs) to improve accuracy and reduce either false accepts or false rejects. The existing research addresses only one of the three significant problems—false predictions—among false predictions, model complexity, and latency. It does not provide an optimized solution to resolve all three issues. This paper proposes a novel wake-word detection model using the Turbo-CRNN architecture with Linear Frequency Cepstral Coefficients (LFCCs) to minimise the adverse effects of the above problems. This paper experimented with multi-head attention, regularizers, and optimizers to determine the best possible model for this application to minimize false positives and negatives. This proposed custom CRNN model achieved higher accuracy, lower model complexity (87k parameters), significantly fewer false predictions, and very low latency (around 80 ms) compared to the CNN, DNN, and TinyCRNN baselines.