Scholarlink

<p>This work offers a useful and effective approach for multilingual speech classification and focuses on the recognition of four Indian languages: English, Bengali, Tamil, and Hindi. Classical language identification techniques typically perform poorly, especially for brief audio samples, because to the varied phonetic features of Indian languages and their limited linguistic resources. A novel system is suggested to effectively tackle this job by leveraging Qdrant, a high-performance vector database for similarity-based classification, and Wav2Vec2, a self-supervised learning model for voice representation. Librosa is used to preprocess voice signals first. Wav2Vec2 then eliminates high-dimensional embeddings. Lastly, techniques for reducing dimensionality maximize computational efficiency. Qdrant stores these embeddings, which enables quick and accurate classification of fresh speech inputs. 85% for English, 90% for Hindi and Tamil, and 95% for Bengali, the suggested model performs well overall with an accuracy of 90%. Comparisons based on accuracy, recall, and F1-score confirm the robustness and dependability of the method. Results indicate that self-supervised learning combined with vector search significantly enhances language recognition, particularly for low-resource languages.</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

Multilingual audio classification: a scalable approach using Wav2Vec2 and qdrant

  • 期刊论文

  • 发表日期：

    2025年12月20日

查看全文
  • Himani Tomar,
  • Deepti Deshwal,
  • Neelu Trivedi

摘要

This work offers a useful and effective approach for multilingual speech classification and focuses on the recognition of four Indian languages: English, Bengali, Tamil, and Hindi. Classical language identification techniques typically perform poorly, especially for brief audio samples, because to the varied phonetic features of Indian languages and their limited linguistic resources. A novel system is suggested to effectively tackle this job by leveraging Qdrant, a high-performance vector database for similarity-based classification, and Wav2Vec2, a self-supervised learning model for voice representation. Librosa is used to preprocess voice signals first. Wav2Vec2 then eliminates high-dimensional embeddings. Lastly, techniques for reducing dimensionality maximize computational efficiency. Qdrant stores these embeddings, which enables quick and accurate classification of fresh speech inputs. 85% for English, 90% for Hindi and Tamil, and 95% for Bengali, the suggested model performs well overall with an accuracy of 90%. Comparisons based on accuracy, recall, and F1-score confirm the robustness and dependability of the method. Results indicate that self-supervised learning combined with vector search significantly enhances language recognition, particularly for low-resource languages.