关键词: 音频检索, 音频指纹, 频率变换

Abstract:

An ideal audio retrieval method can identify all audio clips from large-scale audio database accurately and efficiently. However, none of the methods is robust to all types of distortions. The sampling and counting method based on Philips fingerprint is one of the most efficient methods at present. If it could resist speed changes, including time-stretch and pitch-shift, it would be promising towards an ideal retrieval method. To solve the pitch-shift problem, a sampling and counting method resistant to pitch-shift is proposed, including a method for generating the query fingerprint with changed frequency band intervals, a method for matching the query with multi-scale of pitch and two acceleration strategies, namely, fingerprint extraction in steps and adjustable filtering thresholds. The proposed method can resist pitch-shift from 70% to 130%, and the performance is equivalent to the state-of-the-art QUAD method. Moreover, it can be extended to any retrieval method using Philips-like fingerprint to enhance its ability to resist pitch-shift.

Key words: audio retrieval, audio fingerprint, pitch-shift