Scholarlink

<p>Recent advancements in speech processing have been largely driven by sequence-to-sequence architectures integrated with self-attention mechanisms, with transformers emerging as the dominant paradigm. However, accurately converting spoken language into text remains a significant challenge, especially in linguistically constrained environments, due to the scarcity of labeled speech corpora and the extensive variability in linguistic patterns. The Kashmiri language, in particular, presents unique complexities arising from its diverse regional dialects, phonetic variations, accent disparities, and lack of labeled dataset availability, which significantly impact the performance of automatic transcription models. The introduction of advanced neural architectures, especially transformer-based frameworks, has substantially improved acoustic modeling, with Wav2Vec2 demonstrating remarkable advancements over traditional approaches. While this model effectively reduces reliance on large labeled datasets, its performance deteriorates when applied to speech from underrepresented linguistic groups. In this study, we collected a labeled dataset comprising approximately 10 hours of speech from 14 speakers (eight males and six females). The data were collected from different locations (North Kashmir, South Kashmir, and Central Kashmir) to increase diversity. This study investigates the effectiveness of Wav2Vec2 model variants, specifically XLSR-53 and XLS-R-300 M, and their fine-tuning for speech recognition in the low-resource Kashmiri language. The evaluation results show that XLSR-53 achieved a word error rate (WER) of 47.10% and a character error rate (CER) of 13.53%, while XLS-R-300 M obtained a WER of 48.94% and a CER of 14.23%. The findings indicate that even with limited data and a short training period, the models achieve a relatively low CER, offering valuable insights for automatic speech recognition (ASR) of the Kashmiri language and extending its applicability across diverse linguistic contexts. This study is expected to advance ASR for the Kashmiri language.</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

A Wav2Vec2 model-based automatic speech recognition system for low-resource Kashmiri language

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Muzaffar Ahmad Dar,
  • Jagalingam Pushparaj

摘要

Recent advancements in speech processing have been largely driven by sequence-to-sequence architectures integrated with self-attention mechanisms, with transformers emerging as the dominant paradigm. However, accurately converting spoken language into text remains a significant challenge, especially in linguistically constrained environments, due to the scarcity of labeled speech corpora and the extensive variability in linguistic patterns. The Kashmiri language, in particular, presents unique complexities arising from its diverse regional dialects, phonetic variations, accent disparities, and lack of labeled dataset availability, which significantly impact the performance of automatic transcription models. The introduction of advanced neural architectures, especially transformer-based frameworks, has substantially improved acoustic modeling, with Wav2Vec2 demonstrating remarkable advancements over traditional approaches. While this model effectively reduces reliance on large labeled datasets, its performance deteriorates when applied to speech from underrepresented linguistic groups. In this study, we collected a labeled dataset comprising approximately 10 hours of speech from 14 speakers (eight males and six females). The data were collected from different locations (North Kashmir, South Kashmir, and Central Kashmir) to increase diversity. This study investigates the effectiveness of Wav2Vec2 model variants, specifically XLSR-53 and XLS-R-300 M, and their fine-tuning for speech recognition in the low-resource Kashmiri language. The evaluation results show that XLSR-53 achieved a word error rate (WER) of 47.10% and a character error rate (CER) of 13.53%, while XLS-R-300 M obtained a WER of 48.94% and a CER of 14.23%. The findings indicate that even with limited data and a short training period, the models achieve a relatively low CER, offering valuable insights for automatic speech recognition (ASR) of the Kashmiri language and extending its applicability across diverse linguistic contexts. This study is expected to advance ASR for the Kashmiri language.