Scholarlink

<p>In this study, we propose a supervised method for spoken word recognition designed to perform effectively even with limited training data. The approach begins with audio recordings in the OPUS format, from which text transcripts are extracted using the pre-trained Large XLSR-Wav2Vec2-53 model. Phonemes are then identified from these transcripts using the Carnegie Mellon University (CMU) Pronouncing Dictionary and represented as vectors through LaBSE (Language-Agnostic BERT Sentence Embeddings). Additionally, phoneme bigrams are derived from the transcripts and similarly encoded using LaBSE. Both phoneme and bigram embeddings are processed through a five-layer neural network incorporating batch normalization to enhance learning efficiency. A late fusion strategy is employed to combine the phoneme and bigram representations. The proposed method was evaluated on ten-word categories from the Multilingual Spoken Words Corpus (MSWC), achieving recognition accuracies of 59.38% for Arabic, 24.13% for Vietnamese, and 69.77% for Tamil—surpassing the performance of existing techniques. The findings demonstrate that linguistic features such as phonemes significantly contribute to improving spoken word recognition, particularly under data-scarce or imbalanced conditions. Moreover, results indicate that text-based features derived from pre-trained models outperform traditional audio-based methods. Overall, this work advances Automatic Speech Recognition (ASR) by integrating deeper linguistic representations into the recognition process.</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

PhonoBiEmbedNet: A Phoneme and Bigram Embedding Framework for Low-Resource Spoken Word Recognition

  • 期刊论文

  • 发表日期：

    2026年1月28日

查看全文
  • Sunakshi Mehra,
  • Virender Ranga,
  • Ritu Agarwal

摘要

In this study, we propose a supervised method for spoken word recognition designed to perform effectively even with limited training data. The approach begins with audio recordings in the OPUS format, from which text transcripts are extracted using the pre-trained Large XLSR-Wav2Vec2-53 model. Phonemes are then identified from these transcripts using the Carnegie Mellon University (CMU) Pronouncing Dictionary and represented as vectors through LaBSE (Language-Agnostic BERT Sentence Embeddings). Additionally, phoneme bigrams are derived from the transcripts and similarly encoded using LaBSE. Both phoneme and bigram embeddings are processed through a five-layer neural network incorporating batch normalization to enhance learning efficiency. A late fusion strategy is employed to combine the phoneme and bigram representations. The proposed method was evaluated on ten-word categories from the Multilingual Spoken Words Corpus (MSWC), achieving recognition accuracies of 59.38% for Arabic, 24.13% for Vietnamese, and 69.77% for Tamil—surpassing the performance of existing techniques. The findings demonstrate that linguistic features such as phonemes significantly contribute to improving spoken word recognition, particularly under data-scarce or imbalanced conditions. Moreover, results indicate that text-based features derived from pre-trained models outperform traditional audio-based methods. Overall, this work advances Automatic Speech Recognition (ASR) by integrating deeper linguistic representations into the recognition process.