주파수 변화율을 이용한 음성과 음악의 구분 = Speech and music discrimination using spectral transition rate

다국어 초록 (Multilingual Abstract)

Automatic speech recognition (ASR) is becoming an indispensable technology in many application areas such as telephonic system, ubiquitous system, robot system, and telematics. In real-world environments, ASR faces many kinds of sound sources and they should be discriminated to improve ASR performance. In ASR systems, speech is usually detected from the input signal by voice activity detection (VAD) scheme. Speech and music, however, are not easily discriminated by the VAD because they share similar characteristics such as periodicity and frequency. Speech and music discrimination (SMD) has gained much popularity in recent years for efficient coding and automatic retrieval of multimedia sources and automated speech recognition. Many kinds of approaches have previously been taken to the problem of SMD using spectral energy changes, harmonics, delta cepstral energy, power spectrum deviation, cepstral distance, and etc. However, those feature parameters are not so efficient to get high performance with fast output. The mean of minimum cepstral distance (MMCD) has recently showed high performance. However, the duration used in MMCD was 1 second. Considering the length of a sentence, too much long duration is not proper for the practical system. In this paper, we propose the spectral transition rate (STR) as a novel feature for SMD. We found that the spectral peaks of speech are gradually changing. However, the sound of musical instruments in general tends to keep the peak frequencies and energies unchanged for relatively long period of time compared to speech. The STR is based on this observation. The STR of speech is much higher than that of music. Comparing to other algorithms, the STR based SMD gives relatively fast output without losing its performance. The experimental result shows that the STR based SMD method outperforms a conventional method.

목차 (Table of Contents)

• 제 1 장 서론 1
• 제 2 장 음악과 음성의 주파수 특성 3
• 2.1 음성의 스펙트로그램 3
• 2.2 음악의 스펙트로그램 7
• 2.3 음성과 음악의 주파수 특성 비교 10
• 제 3 장 Spectral Transition Rate 11
• 3.1 STR 11
• 3.2 음성의 STR 13
• 3.3 음악의 STR 16
• 3.4 STR기반의 SMD 19
• 제 4 장 기존의 음성과 음악 검출 알고리즘 22
• 4.1 The mean of minimum cepstral distance 22
• 제 5 장 실험 및 결과 분석 26
• 5.1 실험 환경 26
• 5.2 음악 데이터의 종류 27
• 5.3 실험 결과 27
• 제 6 장 결론 및 향후 과제 38
• 참고문헌 39