Scholarlink

<p>Automatic Speech Recognition (ASR) is an essential component of Human Computer Interaction. With the surge in mobile and home assistant IoT devices, there is a rising need for edge computing solutions supporting ASR applications. Whisper is a state-of-art ASR model that exhibits superior performance. Being resource hungry, this model fails to deliver similar performance on resource constrained devices. Quantization is a thoroughly studied model compression technique in the premise of Deep Learning, which reduces the model size by using smaller integer representations for weights and activations. In this work, quantization is applied to the Whisper releases which are then deployed on Raspberry Pi and Jetson Orin Nano. To get a fair understanding, the models are also deployed on an AWS EC2 instance as well as on a personal computing device. This work is one of the initial attempts to study the performance of quantized Whisper models on edge devices. The quantized models achieved compression ratios of over 3.7 with only a marginal degradation in transcription accuracy. Real-time performance improved significantly, with up to 15% reduction in inference latency observed for the Base model on Raspberry Pi 5. Furthermore, quantized models demonstrated a notable reduction in memory usage, enabling deployment on devices with limited computational resources.</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

Speech recognition in edge environments: an exploration of support and impact of model compression

  • 期刊论文

  • 发表日期：

    2026年1月20日

查看全文
  • Carolene Joy,
  • John Paul Martin,
  • Christina Terese Joseph,
  • Manu Madhavan

摘要

Automatic Speech Recognition (ASR) is an essential component of Human Computer Interaction. With the surge in mobile and home assistant IoT devices, there is a rising need for edge computing solutions supporting ASR applications. Whisper is a state-of-art ASR model that exhibits superior performance. Being resource hungry, this model fails to deliver similar performance on resource constrained devices. Quantization is a thoroughly studied model compression technique in the premise of Deep Learning, which reduces the model size by using smaller integer representations for weights and activations. In this work, quantization is applied to the Whisper releases which are then deployed on Raspberry Pi and Jetson Orin Nano. To get a fair understanding, the models are also deployed on an AWS EC2 instance as well as on a personal computing device. This work is one of the initial attempts to study the performance of quantized Whisper models on edge devices. The quantized models achieved compression ratios of over 3.7 with only a marginal degradation in transcription accuracy. Real-time performance improved significantly, with up to 15% reduction in inference latency observed for the Base model on Raspberry Pi 5. Furthermore, quantized models demonstrated a notable reduction in memory usage, enabling deployment on devices with limited computational resources.