A Speech Recognition Method Based on Domain-Specific Datasets and Confidence Decision Networks.

Area of Science:

• Speech Recognition
• Machine Learning
• Natural Language Processing

Background:

• Automatic Speech Recognition (ASR) systems often struggle with domain-specific vocabulary and out-of-vocabulary (OOV) words.
• Adapting large-scale ASR models to specialized domains requires significant effort and data.

Purpose of the Study:

• To propose a novel speech recognition method enhancing domain adaptability and accuracy.
• To address the challenge of OOV words in realistic ASR scenarios.
• To improve the prediction accuracy of speech models in new domains.

Main Methods:

• Developed a domain-specific language speech network (DSL-Net) and confidence decision network (CD-Net).
• Employed transfer learning using pre-trained model parameters for domain-specific model training.
• Integrated domain-specific models with benchmark models, incorporating importance sampling weights.
• Utilized external knowledge sources to extend the language model and handle OOV words.
• Applied a deep fully convolutional neural network (DFCNN) and Connectionist Temporal Classification (CTC) for domain-specific vocabulary recognition.
• Incorporated a confidence-based classifier to boost overall accuracy and robustness.

Main Results:

• Achieved a significant accuracy improvement from 82% to 91% in the medical domain.
• Domain-specific datasets enhanced performance by 5% to 7% over the baseline.
• Model confidence further improved the baseline by an additional 3% to 5%.

Conclusions:

• Integrating domain-specific datasets and model confidence is crucial for advancing speech recognition technology.
• The proposed DSL-Net and CD-Net approach demonstrates superior adaptability and accuracy in specialized domains.
• This method effectively expands lexical content and improves language model prediction for enhanced ASR performance.