Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Context-Aware Dynamic Chunking for Streaming Tibetan Speech Recognition

Chao Wang, Yuqing Cai, Renzeng Duojie, Jin Zhang, Yutong Liu, Nyima Tashi

TL;DR

The paper tackles low-latency streaming ASR for the morphologically rich Amdo Tibetan by introducing a context-aware dynamic chunking mechanism within a hybrid CTC/Attention framework. It combines a Conformer-based streamable encoder with cross-chunk context propagation and a three-stage training strategy, aligning chunk widths and strides adaptively to speech rate and context. The approach is augmented with linguistically grounded Tibetan lexicons and external LM rescoring, achieving a WER of 6.23% on the test set and demonstrating substantial latency reductions compared with fixed-chunk baselines while approaching global decoding performance. The work delivers a practical streaming solution for Tibetan, supported by a large-scale corpus and resources released to the community, with strong evidence of robustness to long-form speech and varying beam/search configurations.

Abstract

In this work, we propose a streaming speech recognition framework for Amdo Tibetan, built upon a hybrid CTC/Atten-tion architecture with a context-aware dynamic chunking mechanism. The proposed strategy adaptively adjusts chunk widths based on encoding states, enabling flexible receptive fields, cross-chunk information exchange, and robust adaptation to varying speaking rates, thereby alleviating the context truncation problem of fixed-chunk methods. To further capture the linguistic characteristics of Tibetan, we construct a lexicon grounded in its orthographic principles, providing linguistically motivated modeling units. During decoding, an external language model is integrated to enhance semantic consistency and improve recognition of long sentences. Experimental results show that the proposed framework achieves a word error rate (WER) of 6.23% on the test set, yielding a 48.15% relative improvement over the fixed-chunk baseline, while significantly reducing recognition latency and maintaining performance close to global decoding.

Context-Aware Dynamic Chunking for Streaming Tibetan Speech Recognition

TL;DR

The paper tackles low-latency streaming ASR for the morphologically rich Amdo Tibetan by introducing a context-aware dynamic chunking mechanism within a hybrid CTC/Attention framework. It combines a Conformer-based streamable encoder with cross-chunk context propagation and a three-stage training strategy, aligning chunk widths and strides adaptively to speech rate and context. The approach is augmented with linguistically grounded Tibetan lexicons and external LM rescoring, achieving a WER of 6.23% on the test set and demonstrating substantial latency reductions compared with fixed-chunk baselines while approaching global decoding performance. The work delivers a practical streaming solution for Tibetan, supported by a large-scale corpus and resources released to the community, with strong evidence of robustness to long-form speech and varying beam/search configurations.

Abstract

In this work, we propose a streaming speech recognition framework for Amdo Tibetan, built upon a hybrid CTC/Atten-tion architecture with a context-aware dynamic chunking mechanism. The proposed strategy adaptively adjusts chunk widths based on encoding states, enabling flexible receptive fields, cross-chunk information exchange, and robust adaptation to varying speaking rates, thereby alleviating the context truncation problem of fixed-chunk methods. To further capture the linguistic characteristics of Tibetan, we construct a lexicon grounded in its orthographic principles, providing linguistically motivated modeling units. During decoding, an external language model is integrated to enhance semantic consistency and improve recognition of long sentences. Experimental results show that the proposed framework achieves a word error rate (WER) of 6.23% on the test set, yielding a 48.15% relative improvement over the fixed-chunk baseline, while significantly reducing recognition latency and maintaining performance close to global decoding.

Paper Structure

This paper contains 15 sections, 9 equations, 2 figures, 8 tables.

Table of Contents

  1. Introduction
  2. METHODOLOGY
  3. Streamable Encoder-Decoder Architecture with Hybrid CTC/Attention
  4. Context-Aware Dynamic Chunking Mechanism and Training Method
  5. EXPERIMENT
  6. Datasets
  7. Setup
  8. Performance Analysis of Fixed and Dynamic Chunking
  9. Joint Optimization of CTC and Attention Weights
  10. Latency Performance Evaluation
  11. Long-form Speech Generalization Capability
  12. Batch Size and Global Normalization Strategy
  13. Language Model Fusion Effect
  14. CONCLUSION
  15. ACKNOWLEDGEMENT

Figures (2)

  • Figure 1: The overall model architecture is presented, along with a detailed illustration of the attention module specifically enhanced in this work to account for the phonetic characteristics of Tibetan.
  • Figure 2: Structure of the Tibetan word Programming.