JPH05127668A - Automatic transcription device - Google Patents

Abstract

PURPOSE:To provide the automatic transcription device with high precision by taking an analysis so that frequency resolution is high in a low-frequency range and time resolution is high in a high-frequency range. CONSTITUTION:A CPU performs a frequency analyzing process for a music signal, which is sampled and stored in a RAM by an A/D conversion device, by wavelet conversion (S1). Then the output value of the wavelet conversion is inputted and a fundamental frequency component is extracted (S2); and note information such as MIDI codes is generated according to the frequency channel, intensity, start time, and end time of the fundamental frequency component and outputted on a display (S3).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an automatic music transcription device for performing frequency analysis on an acoustic music signal, and more particularly to its frequency analysis processing.

[0002]

2. Description of the Related Art Conventionally, this type of automatic music transcription device has been constructed as shown in FIG. The audio amplifier 1 receives the played music signal as an input and amplifies this signal to an appropriate voltage value. The low-pass filter 2 suppresses aliasing distortion at the time of sampling by passing only the frequency component of 5.5 kHz or less in the signal amplified by the audio amplifier 1. The A / D converter 3 is a low-pass
Sampling frequency of 12 kHz for the filtered signal
z, 16-bit digital signal is converted. I / O port 4 is CPU 5, A / D converter 3, display 8
And are connected. The CPU 5 performs frequency analysis processing of music signal data, extraction processing of basic frequency components, encoding processing, etc., and is connected to the RAM 6 and the ROM 7. The RAM 6 is provided with an area for storing the music signal data sampled by the A / D converter 3 and the frequency analysis result processed by the CPU 5. The RO
M7 stores a frequency analysis logic, a fundamental frequency component extraction logic, an encoding processing logic, and the like. The display 8 displays processing results and the like.

The operation of the conventional example will be described below with reference to FIGS. 2 and 4.

FIG. 2 is a flowchart showing the processing performed by the conventional example.

First, the CPU 5 receives the music signal data s _{i: i = 0, 1, ..., N-1} sampled by the A / D converter 3 and stored in the RAM 6 every 25 msec. Frequency analysis processing is performed by short-time Fourier analysis (S1).

The frequency analysis process (S1) will be described with reference to FIG. First, the CPU 5 initializes a pointer h indicating an observation position for the music signal data (h
= 0: S41). Next, the music signal data stored in the RAM 6 is input, and short-time Fourier analysis is performed. The short-time Fourier analysis is performed by the discrete Fourier transform as shown in equation 1.

[0007]

[Equation 1]

At this time, it is generally known that the frequency resolution of the discrete Fourier transform is uniform in each frequency band and is proportional to L of the equation 1 and the time resolution is inversely proportional to L. In the actual processing, the CPU 5 does not directly calculate the equation 1, but uses the fast Fourier transform to calculate X.
_{k: k = 0,1, ..., L-1} is obtained (S43).

Next, the CPU 5 outputs the logarithmic power spectrum P to the output of the discrete Fourier transform according to the equation (2).
_{k: k-0,1, ..., L-1} are calculated (S44).

[0010]

[Equation 2]

Next, the CPU 5 detects the peak value of the logarithmic power spectrum. Then, the frequency corresponding to the peak value is converted into the scale number of the equal temperament scale based on the general 440 Hz, and the scale number and the peak value are stored in the RAM 6 (S45).

Next, the CPU 5 sets the pointer h to 30
It is incremented by 0 point (25 msec), and the process is returned to (S42) (S46).

Next, the CPU 5 compares the pointer h with the sample number N of the music signal data, and if h <N is "YES", it is explained above (S43 to S46).
When the judgment is “NO”, the frequency analysis processing (S1) is ended (S42).

Next, the CPU 5 inputs the scale number corresponding to the peak of the spectrum stored in the RAM 6 and the peak value thereof, and extracts the fundamental frequency component. As a method of extracting the fundamental frequency component, for example,
The method disclosed in Japanese Patent Laid-Open No. 3-94072 is used (S2).

Next, the CPU 5 creates musical note information such as a MIDI code from the scale number, strength, start time and end time of the fundamental frequency component and outputs the result to the display 8 (S3).

[0016]

SUMMARY OF THE INVENTION In an automatic transcription apparatus, it is an object to process a music signal and identify a fundamental frequency of a musical instrument sound. A typical equal tempered scale is an octave interval logarithmically divided into 12 equal parts, and a frequency analysis process requires a frequency resolution of 1/12 octave. That is, high frequency resolution is required in the low frequency region, but not very high frequency resolution in the high frequency region. However, as a characteristic of music, generally, a high time resolution is often required in a high frequency region.

However, in the frequency analysis processing of the above-mentioned conventional automatic transcription apparatus, short-time spectrum analysis by discrete Fourier transform as shown in equation 1 is used, but the frequency resolution of discrete Fourier transform is several as described above. It is known that it is proportional to L of 1 and the temporal resolution is inversely proportional to L, and they are uniform in each frequency band. Therefore,
When trying to increase the frequency resolution in the low frequency region, the time resolution in the high frequency region becomes insufficient. On the other hand, there is a drawback in that the frequency resolution in the low frequency region becomes insufficient when trying to increase the time resolution in the high frequency region.

The present invention solves the above-mentioned problems, and by using an analysis method in which the frequency resolution is high in the low frequency region and the time resolution is high in the high frequency region, a highly accurate automatic transcription device is provided. Is intended to provide.

[0019]

To achieve this object, the present invention provides a frequency analysis means for performing frequency analysis with acoustic signal data as input, and a fundamental frequency component for extracting a fundamental frequency component from the frequency analysis result. From the extraction means and the pitch, strength, start time, end time, etc. of the fundamental frequency component, MID
An encoding means for outputting note information such as an I-code,
The frequency analysis means is constructed by wavelet transform.

Further, the center frequency of the output of the wavelet transform may be set to be equal to each frequency of the equal tempered scale based on the performance altitude.

A Gabor function may be used as the basic wavelet function of the wavelet transform.

[0022]

The frequency analyzing means of the present invention having the above-mentioned structure receives the acoustic signal as an input and performs the frequency analysis by the wavelet transform. The fundamental frequency component extracting means extracts the fundamental frequency component from the frequency analysis result. The encoding means outputs note information such as MIDI code from the pitch, strength, start time, end time, etc. of the fundamental frequency component.

[0023]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of an automatic music transcription device according to this embodiment. The audio amplifier 1 constituting the present embodiment receives the played music signal as an input and amplifies this signal to an appropriate voltage value. The low-pass filter 2 is 5.5 kHz in the signal amplified by the audio amplifier 1.
The aliasing distortion at the time of sampling is suppressed by passing only the frequency components equal to or lower than z. The A / D converter 3 converts the low-pass filtered signal into the sampling frequency 12
Converted to a 16-bit digital signal of kHz. I /
The O port 4 connects the CPU 5, the A / D conversion device 3, and the display 8. The CPU 5 performs frequency analysis processing of music signal data, extraction processing of basic frequency components, encoding processing, etc., and is connected to the RAM 6 and the ROM 7. The RAM 6 is provided with an area for storing the music signal data sampled by the A / D converter 3 and the frequency analysis result processed by the CPU 5. The ROM 7 stores a frequency analysis logic, a fundamental frequency extraction logic, an encoding processing logic and the like. The display 8 displays processing results and the like.

The operation of this embodiment will be described below with reference to FIGS.
Will be described.

FIG. 2 is a flow chart showing the processing performed by this embodiment.

First, the CPU 5 performs frequency conversion by wavelet transformation on the music signal data s _{i: i = 0,1, ..., N-1} sampled by the A / D converter 3 and stored in the RAM 6. Analysis processing is performed (S1).

Here, the wavelet transform will be described. Wavelet transform F _{(a, b) of} one-dimensional signal f _{(t )}
Is defined by Equation 3.

[0029]

[Equation 3]

Here, Ψ is a function called a basic wavelet function. The wavelet transform is defined as an inner product of a parameter a representing a scale with respect to frequency and a family of functions created from Ψ by moving a parameter b representing time and a function f _(t) to be analyzed. Ψ ^* is the complex conjugate of Ψ. In this embodiment, a, b
Is defined by Equations 4 and 5, and the digitized input signal s
_The discretized wavelet transform S _{m, n} for _i is given by
Define in.

[0031]

[Equation 4]

[0032]

[Equation 5]

[0033]

[Equation 6]

Here, 1 / T _s is the sampling frequency (12000 in this embodiment). m and n are parameters relating to frequency and observation position (time), respectively, and will be referred to as frequency channel number and observation position in the present specification. The smaller the frequency channel number m, the higher the frequency range, and the larger the frequency channel number m, the lower the frequency range. Further, as the basic wavelet function Ψ _(t) , a Laplacian-Gaussian function, one using an auditory system impulse response, etc. are known, but in this embodiment, uncertainty about time and frequency is minimum, and in this sense, The Gabor function, which is said to have the best localization in the time-frequency space, is used. The Gabor function is defined by Equation 7.

[0035]

[Equation 7]

From Equations 6 and 7, it is understood that the time resolution is inversely proportional to a, that is, the smaller a (high frequency region), the higher the time resolution, and the larger a (low frequency region), the lower. In addition, the Fourier transform Ψ ₍ ω ₎ of the Gabor function is as shown in ^Equation 8, and the Fourier transform of a ⁻¹ / ² · Ψ _{((tb) / a)} is a ^1/2 · Ψ ₍ aω ₎ · exp [− iωb], the frequency resolution is proportional to a, and it can be seen that the larger a is (low frequency region), the higher the frequency resolution, and the smaller a is (high frequency region), the lower.

[0037]

[Equation 8]

As described above, by calculating equation 6, it is possible to perform analysis such that the frequency resolution is high in the low frequency region and the time resolution is high in the high frequency region.

In Expression 7, ω _p is the output center frequency for the highest frequency channel (m = 0) multiplied by 2π, and in this embodiment ω _p = 2π × 5586H
z. As a result, the center frequency of the output of each frequency channel can be made approximately equal to the frequency of the equal temperament scale based on the general 440 Hz. Further, the half width in the frequency space is given by (2ω _p / r), and in this embodiment, r = 30.

In the present embodiment, M = 72. This enables analysis of the frequency band from 87 Hz to 5586 Hz. N is the number of samples of the music signal data stored in the RAM 6.

Here, the frequency analysis process (S
The operation of 1) will be described. First, the CPU 5 initializes the frequency channel number m (m = 0: S3).
1). Next, the CPU 5 initializes the observation position n (n
= 0: S33). Next, the CPU 5 uses S by the above equation 6.
_{m and n} are calculated (S35). Next, the CPU 5 logarithmizes S _{m, n} by the equation 8 and stores it in the RAM (S36).

[0042]

[Equation 9]

Next, the CPU 5 increments the observation position n and returns the processing to (S34) (S37). CPU5
Compares the observation position n with the sample number N of the music signal data, and if the judgment of n <N is “YES”, it is explained above (S
The processes of 35 to S37 are repeated, and if “NO”, the process proceeds to (S38) (S34). When the determination is (NO) in (S34), the CPU 5 increments the frequency channel number m and returns the process to (S32) (S3).
8). The CPU 5 compares the frequency channel numbers m and M, and repeats the above-described processing of (S33 to S38) if the determination of m ≦ M is “YES”, and ends the frequency analysis processing (S1) if “NO”. Yes (S32).

Next, the CPU 5 receives the output value of the wavelet transform stored in the RAM 6 as an input, and extracts the fundamental frequency component. As a method of extracting the fundamental frequency component, for example, the method disclosed in Japanese Patent Laid-Open No. 3-94072 is used (S2).

Next, the CPU 5 determines M from the frequency channel number, strength, start time, end time, etc. of the basic frequency component.
Note information such as an IDI code is created and output to the display 8 (S3).

The present invention is not limited to the configuration described in detail above, and various modifications can be made without departing from the spirit of the invention.

[0047]

As is apparent from the above description, according to the automatic music transcription apparatus of the present invention, the frequency analysis means for performing frequency analysis with an acoustic signal as an input, and the fundamental frequency component extracted from the frequency analysis result. In the automatic music transcription device, the frequency analysis means is provided with a fundamental frequency component extracting means and a coding means for outputting musical note information such as MIDI code from the pitch, strength, start time, end time, etc. of the fundamental frequency component. By implementing the wavelet transform, it is possible to perform frequency analysis suitable for music notation processing, which requires high frequency resolution in the low frequency region and high time resolution in the high frequency region.

[Brief description of drawings]

FIG. 1 is a block diagram according to an embodiment of the present invention.

FIG. 2 is a flowchart of an automatic music transcription device according to an embodiment of the present invention.

FIG. 3 is a flowchart of frequency analysis processing.

FIG. 4 is a flowchart of a conventional frequency analysis process.

[Explanation of symbols]

 1 Audio amplifier 2 Low pass filter 3 A / D converter 5 CPU 8 Display

Claims (3)

[Claims] 1. A frequency analysis means for inputting acoustic signal data to perform a frequency analysis, a fundamental frequency component extraction means for extracting a fundamental frequency component from the frequency analysis result, a pitch, a strength, and a start time of the fundamental frequency component. , From the end time, etc. to MI
An automatic music transcription device comprising an encoding means for outputting musical note information such as a DI code, wherein the frequency analysis means is constituted by a wavelet transform. 2. The automatic music transcription device according to claim 1, wherein the center frequency of the output of the wavelet transform is equal to each frequency of the equal tempered scale with the performance altitude as a reference. 3. The automatic music transcription device according to claim 1, wherein the basic wavelet function of the wavelet transform is a Gabor function.