Comparison and Optimization of Amdo Tibetan Speech Recognition Based on Transducer

Amdo Tibetan, as a low-resource language with a complex phonological system, faces dual challenges in speech recognition: data scarcity and insufficient model generalization. Under the condition of limited Tibetan corpus resources, this study constructs a Transducer model based on the Kaldi-2 framework. To address the complex syllable structures (e.g., consonant clusters) and tonal variations in Tibetan, we introduce noise injection during data processing, integrate Tibetan n-gram language models, and employ BPE subword tokenization to optimize training data. We compare the performance of two encoders, Conformer<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[1]</sup> and Zipformer<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">[2]</sup>, under seven decoding strategies in low-data scenarios. Experimental results demonstrate Zipformer’s superiority in multi-scale temporal modeling. The optimized Zipformer outperforms Conformer across multiple dimensions: under the Fast Beam Search N-Best Language Model Enhanced decoding rule, Zipformer reduces the word error rate (WER) from $11.72 \%$ to $6.44 \%$, achieves a $22.6 \%$ improvement in inference speed, and reduces memory consumption to 0.9 GB. These results highlight Zipformer’s efficiency and accuracy in handling complex speech recognition tasks, attributed to its multi-scale modeling capability and adaptability to long-term dependencies in Tibetan. This research provides practical guidelines for model selection and optimization in low-resource, phonologically complex languages, offering a feasible technical solution for digitizing minority languages.