Dynamic Acoustic Unit Augmentation with BPE-Dropout for Low-Resource End-to-End Speech Recognition

Area of Science:

• Speech Technology
• Machine Learning
• Natural Language Processing

Background:

• On-device automatic speech recognition (ASR) is crucial for speech assistants, with end-to-end models preferred for efficiency and quality.
• End-to-end ASR requires substantial speech data and struggles with personalization, particularly out-of-vocabulary (OOV) words.
• Low-resource scenarios with high OOV rates present significant challenges for developing effective ASR systems.

Purpose of the Study:

• To develop an effective end-to-end ASR system for low-resource conditions with high OOV rates.
• To address the challenge of recognizing OOV words in personalized ASR.
• To improve ASR performance without increasing computational demands.

Main Methods:

• Proposed a dynamic acoustic unit augmentation method using Byte Pair Encoding with dropout (BPE-dropout).
• BPE-dropout non-deterministically tokenizes utterances to enhance token context and regularize distributions.
• This method reduces the need for searching optimal subword vocabulary sizes.

Main Results:

• Achieved steady improvements in both regular and personalized (OOV-focused) ASR tasks.
• Demonstrated at least a 6% relative reduction in word error rate (WER) and a 25% relative increase in F-score.
• A monolingual Turkish Conformer model using BPE-dropout reached competitive results (22.2% CER, 38.9% WER).

Conclusions:

• BPE-dropout effectively enhances end-to-end ASR in low-resource, high-OOV environments.
• The technique offers significant performance gains without additional computational cost.
• The proposed method shows promise for improving ASR systems for personalized and resource-constrained applications.