基于机器学习的基频估计算法架构  

Machine Learning Based Algorithm Architecture for Base Frequency Estimation

作　　者：陈海波 彭成伟 CHEN Haibo;PENG Chengwei(Institute of shootwave communication,Handa Technology co.,ltd of Nanjing panda co.,ltd.,Jiangsu Nanjing 210007;Command and Control Engineering College,Army Engineering University of PLA,Jiangsu Nanjing 210001)

机构地区：[1]南京熊猫汉达科技有限公司短波通信技术研究所,江苏南京210007 [2]中国人民解放军陆军工程大学指挥控制工程学院,江苏南京210001

出　　处：《长江信息通信》2025年第1期73-75,共3页Changjiang Information & Communications

摘　　要：基音频率是声音信号的重要特征之一,反映了声音的情感、说话者身份和音色特征,广泛应用于语音合成、音乐合成、语音转换等领域。传统的基音频率估计方法主要依赖信号处理技术,基于时域和频域特征进行分析和提取。采用传统信号处理方法与新兴的机器学习方法进行比较,研究了如何提高复杂语音信号中基音频率估计的准确性。传统方法,如Yin算法,侧重于时域分析,而机器学习方法,如CREPE,利用大量训练数据学习基音模式。结果显示,尽管传统方法易于实现,但在处理复杂信号时准确性不足。相比之下,机器学习方法在复杂语音和音乐信号的基音估计中表现出更高的准确性和鲁棒性。结论指出,随着技术的进步,机器学习方法正逐渐取代传统方法,成为基音频率估计的主流技术。Fundamental frequency(pitch)is a crucial characteristic of sound signals,reflecting aspects such as emotion,speaker identity,and timbre.It is widely applied in fields like speech synthesis,music synthesis,and voice conversion.Traditional pitch estimation methods mainly rely on signal processing techniques,analyzing and extracting features based on time and frequency domains.A comparison between traditional signal processing methods and emerging machine learning methods is investigated to improve the accuracy of fundamental frequency estimation in complex speech signals.Traditional methods,such as the Yin algorithm,focus on time-domain analysis,while machine learning methods,such as CREPE,utilize large amounts of training data to learn the fundamental tone patterns.The results show that although the traditional methods are easy to implement,they are not accurate enough when dealing with complex signals.In contrast,machine learning methods show higher accuracy and robustness in fundamental tone estimation for complex speech and music signals.The conclusion states that with the advancement of technology,machine learning methods are gradually replacing traditional methods as the mainstream technique for fundamental frequency estimation.

关 键 词：基音频率估计 机器学习 基频估计算法 

分 类 号：TN391[电子电信—物理电子学]