2007 Multimedia System Final Paper Presentation

Music Recognition 492410021 蘇冠年

492410070 蔡尚穎

Introduction

• In future, the problem is not anymore how to get access to multimedia content, the task is how to find what I’m looking for…

Music Recognition System

Training

Data Base

Recognition

Result

Input Data

Before the Algorithm

• Practical Problems

- Disturbance of noise

- Disturbance of Harmonic

- Singer and instrument

- …

Algorithm I

• Pitch detection - notes, chords …

• Based on frequency domain

- according to music characteristics, it analyzed spectrum at the music pitches

- using Wavelet Transform and DTFT (Discrete-Time Fourier Transform)

Frequency Analysis

• Music signal is of typical time-frequency distribution

and has short-time steady property

Frequency Analysis

• Wavelet Transform

- Daub4 Wavelet base by Mallet Algorithm

• DTFT to calculate amplitude

- pitch frequency as parameter ω

Frequency Analysis

• Analyzed result

Notes Recognition

• Step 1: Note Voting - 1. analyzed each data by wavelet transform in frequency domain.

- 2. picked out a numbers of notes that have biggest amplitudes in a data as candidate notes.

- 3. count of the appearance times of the candidate notes in several neighbor dada

• Step 2 : denote the different segments of the music

• Step 3 : selected the note that appears most and has the biggest average amplitude

- A piece of music

- Wave form of the data

- the spectrogram of segment 1

- determine the note

Chords Recognition

• What is the chord ?

• The chord components always have the similar amplitude in frequency domain

Chords Recognition

• Step 1 : define as a set of candidate notes

and as the amplitude of the notes p

• Step 2 : calculate likelihood coefficient of each note

• Step 3 : coefficient L is the average likelihood coefficient among the notes in a candidate chord

- A piece of music

- Wave form of the data

- the spectrogram of segment 1

- determine the chords

Algorithm II

• Items of recognizing

• Single-pitched melody

• Multiple-instrument melody

Pre-Processing

Adaptive Template-matching

• Phase Tracking

• Template Filtering

Phase Tracking

z : input signal

r , i : possible sound

p : narrow-band filter

Phase Tracking

• fs : sampling frequency

• fc : center frequency of the band-pass filter

Template Filtering

• minimization of J

z(k) : input sum of template waveforms

hn(m) : convolution of the filter coefficients

rn(k) : phase-adjusted waveform

•

Template Filtering

Music Stream Networks

• Problem of local information

• Bayesian probabilistic network

Conclusion

Reference

1. Zheng Cao, Shengxiao Guan, Zengfu Wang. “A Real-time Algorithm for Music Recognition Based on Wavelet Transform” IEEE June 21 - 23, 2006, Dalian, China

2. Kunio Kashino ,Hiroshi Murase . “Music Recognition using Note Transition Context”

IEEE 1998, NTT Basic Research Laboratories

3. Karlheinz Brandenburg. “Digital Entertainment: Media technologies for the future”

IEEE 2006 , Fraunhofer IDMT & Technische Universität Ilmenau

4. Chen Genfand, Xia Shunren. “The study and prototype system of printed music recognition”. IEEE 2003

5. D Bainbridge , T C Bell. “Dealing with superimposed objects in optical music recognition” IEEE 15-17 July 1997 Universities of Waikato and Canterbury, New Zealand

6. MALLAT'S FAST WAVELET ALGORITHM: RECURSIVE COMPUTATION OF

CONTINUOUS-TIME WAVELET COEFFICIENTS