一种提高ASR识别效果的人工干预分词方法

By constructing a special word segmentation rule library and combining it with the SentencePiece word segmentation tool, a word segmentation algorithm was designed. Utilizing the Transformer and CTC loss function architecture model, the problem of difficulty in selecting word sub-word granularity in English speech recognition was solved, achieving more efficient and accurate speech recognition.

CN120833785BActive Publication Date: 2026-07-17读书郎教育科技有限公司
View PDF 3 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
读书郎教育科技有限公司
Filing Date
2025-08-18
Publication Date
2026-07-17

Smart Images

  • Figure CN120833785B_ABST
    Figure CN120833785B_ABST
Patent Text Reader

Abstract

本发明涉及人工智能英文语音识别领域，尤其涉及一种提高ASR识别效果的人工干预分词方法。本发明通过构建特殊分词规则库，基于该库设计分词算法，对数据文本进行处理以得到token序列及id数组，同时提取音频频谱特征，将两者转化为数学向量。利用音频特征向量训练初始化的Transformer与CTC损失函数架构模型，多次迭代优化参数。当训练损失和验证损失稳定且不再显著下降时，得到最优模型参数。该方法有效解决了传统分词方法中粒度选择的难题，避免过度切分或切分粒度过大导致词表token过大、训练难度增加的问题，从而增强分词合理性，遵循语言规律地分词方式，提高了语音识别的准确率与高效性。
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Speech recognition model training method, speech recognition method and related devices

    CN112017643A

  • Text matching method and device

    CN111143515A

  • Vehicle-mounted voice understanding method and system based on vocabulary enhancement and multi-task structure

    CN118197283A