Scholarlink

<p>The advent of large-scale machine learning models has transformed artificial intelligence, particularly in natural language processing, by enabling efficient fine-tuning of pre-trained models. Nevertheless, fully fine-tuning these models remains computationally intensive, as it involves updating all parameters. To mitigate this challenge, Low-Rank Adaptation (LoRA) techniques have been introduced, offering a parameter-efficient alternative that reduces computational demands without compromising performance. This work focuses on the automatic speech recognition of Tarifit, a low-resource dialect of the Amazigh language spoken in Morocco. Tarifit is characterized by its variability across different regions, further complicating speech recognition tasks. We collected a dataset of 11,644 audio samples, comprising 200 isolated words and 100 sentences recorded by native speakers from the Rif region. To address the limited data and the speech variability inherent in Tarifit, we propose a few-shot learning paradigm that utilizes a pre-trained Whisper model. The model learns generic feature embeddings through its encoder, which are paired with a classification head for speech recognition. The fine-tuning process leverages LoRA and three of its variants (DoRA, AdaLoRA, and QLoRA) to enhance efficiency and minimize computational costs. The experimental results validate the success of this approach, with our best models using LoRA and DoRA, both reaching an accuracy of 98.91% while only around 6% of the parameters were trainable. AdaLoRA achieved 96.85%. Moreover, QLoRA further optimized the process, reducing fine-tuning duration by 52.64% while maintaining a competitive accuracy of 98.22%.</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

Efficient fine-tuning of whisper models for Amazigh ASR: leveraging LoRA and its variants

  • 期刊论文

  • 发表日期：

    2026年2月25日

查看全文
  • Mohamed Daouad,
  • Fadoua Ataa Allah,
  • El Wardani Dadi

摘要

The advent of large-scale machine learning models has transformed artificial intelligence, particularly in natural language processing, by enabling efficient fine-tuning of pre-trained models. Nevertheless, fully fine-tuning these models remains computationally intensive, as it involves updating all parameters. To mitigate this challenge, Low-Rank Adaptation (LoRA) techniques have been introduced, offering a parameter-efficient alternative that reduces computational demands without compromising performance. This work focuses on the automatic speech recognition of Tarifit, a low-resource dialect of the Amazigh language spoken in Morocco. Tarifit is characterized by its variability across different regions, further complicating speech recognition tasks. We collected a dataset of 11,644 audio samples, comprising 200 isolated words and 100 sentences recorded by native speakers from the Rif region. To address the limited data and the speech variability inherent in Tarifit, we propose a few-shot learning paradigm that utilizes a pre-trained Whisper model. The model learns generic feature embeddings through its encoder, which are paired with a classification head for speech recognition. The fine-tuning process leverages LoRA and three of its variants (DoRA, AdaLoRA, and QLoRA) to enhance efficiency and minimize computational costs. The experimental results validate the success of this approach, with our best models using LoRA and DoRA, both reaching an accuracy of 98.91% while only around 6% of the parameters were trainable. AdaLoRA achieved 96.85%. Moreover, QLoRA further optimized the process, reducing fine-tuning duration by 52.64% while maintaining a competitive accuracy of 98.22%.