基于内容的音频检索技术研究与实现

【作者】 邢伟利；

【导师】 周明全；

【作者基本信息】 西北大学 ， 计算机软件与理论， 2004， 硕士

【摘要】 如何有效地对海量数据尤其是诸如音频之类的多媒体数据进行分析、存储和检索是一个亟待解决的问题。由于原始音频数据的非结构化特性，音频检索受到极大的限制。相对于日益成熟的图像与视频检索，音频检索相对滞后。基于内容的音频检索已成为多媒体检索技术的研究热点。本文对基于内容的音频检索的关键技术展开分析，主要在以下几个方面开展了工作： 1、音频信号特征提取与表达。音频检索是多特征组合检索，对音频信号的感知特征如响度、亮度、音调等；物理特征如过零率、Mel倒谱系数、线性预测系数等进行了分析，不同的特征组合应用于不同类型的音频检索。 2、音频分割与识别。将音频分层分割算法改进为基于模板的音频分割算法，利用隐马尔可夫模型的良好的随机时序性以及不依赖于具体的应用阈值的优势，使分割识别准确率有较大提高。随着MPEG压缩格式成为多媒体编码主流，研究了直接对MP3格式的音频信号提取特征，基于MPEG压缩域特征音频分割。 3、基于内容的音频检索。从音频示例查询的角度，针对不同的音频例子表示方法，分别研究了基于隐马尔可夫模型分类模板的音频例子检索算法，和基于模糊聚类音频例子检索算法。针对音乐(歌曲)检索独有特性，研究了基于哼唱的音乐(歌曲)检索，实验证明，此算法有一定的准确性。 本文设计实现了一个基于内容的音频检索原型系统，是具有良好扩展性的系统，实现了高速、有效的音频检索。最后对基于内容的音频检索系统的发展趋势研究热点进行了展望。更多还原

【Abstract】 How to analyze . store and retrieve the huge amount of data efficiently and effectively, especially forthose multimedia data is an imperative problem. Because of the nonstructural character, the development of audio retrieval is restricted hugely. Comparing with image and video retrieval, audio retrieval study is behindhand. Content-based audio retrieval has been the studied hotspot of multimedia retrieval. This paper focuses on the key techniques of content-based audio retrieval, developed mainly in the following aspect:1. This paper analyzes audio feature extraction and expression. Audio retrieval completes through multiple features combination. This paper studies the audio perceptive features, such as loudness, brightness and pitch etc, and the audio physical feature, such as zero-crossing rate, linear prediction coefficient and Mel cepstrum coefficient etc. Different features combination can be applied in different audio retrieval.2. This paper study audio segmentation and recognize, and proved the audio layered segmentation algorithm to template-based audio segmentation algorithm, making use of the Hidden Markov Model’s better stochastic sequence and superiority independent of concrete threshold, the veracity of segmentation and recognize is enhanced a lot Because compressed audio format MPEG has been the mainstream of multimedia encoding, this paper studies the feature extraction onMP3 directly and audio segmentation on MPEG.3. This paper study content-based audio retrieval. Studying from query by audio example, aiming at different audio example expression, mis paper studies the query by audio example algorithm based on Hidden Markov Model classification template and fuzzy clustering centroid respectively Aiming at music(song)’s unique character, mis paper studies muskXsong) retrieval by Humming, and does some performance test, the algorithm has high veracity.更多还原