基于自适应阈值与基频检测的自发性口语音频分割算法

SPONTANEOUS ORAL SPEAKING AUDIO SEGMENTATION ALGORITHM BASED ON ADAPTIVE THRESHOLD AND PITCH DETECTION

为了去除自发性口语音频中静音和噪音段的干扰,提高语音识别率和解码识别效率,提出一种音频能量自适应阈值计算方法。针对实时自动口语评测应用,设计了能量阈值自适应系数,该方法将根据能量阈值自适应系数动态地给每个考生的个人单次所有考试音频计算匹配一个能量阈值,以避免阈值选择和硬门限判决造成的误检。在基于自适应能量阀值的音频切分后,加入了基频检测步骤,以判别切分后所得音频段是否为噪声,从而最终分离出纯净的口语音频部分。实验结果表明,该算法能有效准确地切分音频,且鲁棒性较强。

We present an audio energy adaptive threshold calculation method in order to remove the interference of silent and noisy segments in spontaneous oral speaking audio and to improve speech recognition rate and decoding efficiency.Aiming at the application of real-time automatic oral speaking evaluation,we design the energy threshold adaptive coefficient.This method will dynamically calculate and match an energy threshold to all personal single examining audios for every examinee based on the energy threshold adaptive coefficient in order to avoid the detection errors due to threshold selection and hard threshold judging.The pitch detection procedure is added after the audio segmentation based on adaptive energy threshold for estimating whether the segmented audio segments are noises,so that the pure audio components of oral speaking are separated finally.Experimental results show that the proposed algorithm can effectively segment audio,and is quite robust as well.