Improving Burmese Clinical Speech Recognition

Summary & Methodology Analysis

The team addressed the lack of specialized Burmese medical speech data by building the myMediWhisper corpus. This involved manually verifying Burmese translations of clinical dialogues and recording them at 16 kHz in controlled environments. To simulate real-world audio challenges, they constructed training data using the Pyroomacoustics library to model various room impulse responses and applied multi-stage data augmentation. This process included waveform-level techniques such as pitch shifting and additive Gaussian noise, alongside spectrogram-level masking methods to improve model generalization. Fine-tuning, the process of adjusting a pre-trained model on domain-specific data, was performed using either full fine-tuning or parameter-efficient fine-tuning (PEFT) via Rank-Stabilized LoRA (rsLoRA). rsLoRA is a technique for updating only a small subset of model weights to reduce training overhead while maintaining performance. The resulting myMediWhisper-Medium model achieved a state-of-the-art word error rate of 23.44% on baseline tests without augmentation. While data augmentation helped in noisy and reverberant conditions, it showed reduced performance on clean speech samples. The authors note limitations regarding speaker diversity and regional dialect representation within the corpus. Furthermore, because robustness was evaluated using synthetic data, the performance in actual clinical environments may vary compared to the reported metrics. The paper does not specify the training time, hardware specifications, or absolute latency figures for the models.