Scholarlink

<p>Developing robust speech technologies for a low-resource tribal language like Kui is challenging due to the unavailability of large scale annotated audio data and other linguistic resources. This paper presents investigation into Automatic Speech Recognition (ASR) and Speaker Identification (SID) for the Kui language. A recently developed corpus has been used for both the experiment. Here, the pre-trained IndicWav2Vec2 model that incorporates a weighted Connectionist Temporal Classification (CTC) loss to address the grapheme frequency imbalance is compared with the standard CTC for ASR. The Character Error Rate (CER) and Word Error Rate (WER) metrics are used to evaluate the performance of the fine-tuned models. In addition to this, to enhance generalization, speaker diversity is introduced in the training set to develop models whose performance is compared with the training set of one speaker model. In parallel, the study also explores and compares various deep-learning architectures for speaker identification in Kui. It is basically a closed set speaker identification task. The different architectures that are focused on here are grouped into three distinct categories as Transformer-based, Embedding-based, and Pooling-based. For Transformer-based models, pre-trained large-scale speech models like Wav2Vec2 and HuBERT with partial fine-tuning for speaker identification are considered. Along with the standard accuracy measure, McNemar’s test is also used to find statistical significance of the models. For the embedding-based approach, ECAPA_TDNN network is utilized to extract speaker embeddings. These embeddings are then passed to different classifiers such as Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), and Random Forest (RF). To investigate pooling-based methods, techniques such as Average Statistical Pooling (ASP), Self-Attentive Pooling (SAP), and other statistical pooling mechanisms were integrated with Time-Delay Neural Network (TDNN) features. These were followed by fully connected layers with softmax activation. Experimental results demonstrate that the Transformer-based models achieve the highest accuracy, indicating its effectiveness in capturing relevant speaker-specific information.</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

Exploring speech technologies for the Kui language in a low-resource scenario

  • 期刊论文

  • 发表日期：

    2026年7月7日

查看全文
  • Sanjibani Sudha Pattanayak,
  • Ajit Kumar Nayak,
  • SmitaPrava Mishra,
  • Prithviraj Mohanty,
  • Sonali Das,
  • Aum Auroansu Meher

摘要

Developing robust speech technologies for a low-resource tribal language like Kui is challenging due to the unavailability of large scale annotated audio data and other linguistic resources. This paper presents investigation into Automatic Speech Recognition (ASR) and Speaker Identification (SID) for the Kui language. A recently developed corpus has been used for both the experiment. Here, the pre-trained IndicWav2Vec2 model that incorporates a weighted Connectionist Temporal Classification (CTC) loss to address the grapheme frequency imbalance is compared with the standard CTC for ASR. The Character Error Rate (CER) and Word Error Rate (WER) metrics are used to evaluate the performance of the fine-tuned models. In addition to this, to enhance generalization, speaker diversity is introduced in the training set to develop models whose performance is compared with the training set of one speaker model. In parallel, the study also explores and compares various deep-learning architectures for speaker identification in Kui. It is basically a closed set speaker identification task. The different architectures that are focused on here are grouped into three distinct categories as Transformer-based, Embedding-based, and Pooling-based. For Transformer-based models, pre-trained large-scale speech models like Wav2Vec2 and HuBERT with partial fine-tuning for speaker identification are considered. Along with the standard accuracy measure, McNemar’s test is also used to find statistical significance of the models. For the embedding-based approach, ECAPA_TDNN network is utilized to extract speaker embeddings. These embeddings are then passed to different classifiers such as Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), and Random Forest (RF). To investigate pooling-based methods, techniques such as Average Statistical Pooling (ASP), Self-Attentive Pooling (SAP), and other statistical pooling mechanisms were integrated with Time-Delay Neural Network (TDNN) features. These were followed by fully connected layers with softmax activation. Experimental results demonstrate that the Transformer-based models achieve the highest accuracy, indicating its effectiveness in capturing relevant speaker-specific information.