The challenge of recognizing dysarthric speech, particularly among children, stems from pronounced acoustic variability due to impaired articulatory precision. Past research has demonstrated improved recognition through hybrid DNN/HMM sequence discriminative training. A new paper by Sapkota, Paban, Kathania, Hemant Kumar, Kurimo, Mikko, Kadiri, Sudarsana Reddy, and Narayanan, Shrikanth, titled "Cross-Dataset, Age, and Gender Generalization: A Comprehensive Analysis of Fine-Tuning Strategies for Low-Resource Children's ASR," presents a thorough investigation of various combinations of acoustic features tailored to different acoustic models, offering suitable feature selections for each.
Key Findings
The researchers systematically examined the TORGO database, which is commonly used for dysarthric speech research. They demonstrated the potential to enhance the performance of the state-of-the-art Factorized Time Delay Neural Network (F-TDNN) model for recognizing dysarthric speech. Their methods yielded a 4.65% relative improvement in isolated word recognition and a 4.63% relative improvement in sentence recognition compared to previous research, effectively compensating for speech variability.
| Metric | Relative Improvement | Model Used |
|---|---|---|
| Isolated word recognition | 4.65% | F-TDNN |
| Sentence recognition | 4.63% | F-TDNN |
The improvement is attributed to the deliberate selection of the number of overlapping frames between consecutive training example chunks, which helped address acoustic variability.
Methodology
The study explored various acoustic features, with the incorporation of Pitch features notably improving recognition performance, especially for sentence recognition tasks involving dysarthric speech. The research builds on earlier work using hybrid DNN/HMM sequence discriminative training, which has been effective in improving recognition accuracy. By focusing on the F-TDNN architecture, the authors optimized the fine-tuning process for low-resource children's ASR.
Implications for Low-Resource ASR
This research addresses a critical gap in speech recognition technology: the generalization across different datasets, ages, and genders for children's speech, particularly when resources are limited. The findings suggest that targeted fine-tuning strategies, such as using pitch features and adjusting frame overlap, can significantly boost performance without requiring large amounts of training data. This could enable more robust ASR systems for children with speech impairments in various real-world applications.
