\n",
"In this notebook, we discuss temporal smoothing and downsampling techniques for postprocessing feature representations. Parts of the notebook follow Section 3.1.2.3 and Section 7.2.1 of [Müller, FMP, Springer 2015].\n",
" \n",
"
\n",
"\n",
" "
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Beethoven Example\n",
"\n",
"There are many ways of converting music recordings into chroma-based feature representations. For example, we studied how to derive [chroma features from a spectrogram](../C3/C3S1_SpecLogFreq-Chromagram.html) by suitably pooling Fourier coefficients. By applying pre- and postprocessing steps, the properties of chroma features can be changed significantly. For example, we have seen that one can increase the robustness to variations in timbre or sound intensity by performing additional [logarithmic compression](../C3/C3S1_LogCompression.html) and [normalization](../C3/C3S1_FeatureNormalization.html) steps. In this notebook, we study further postprocessing techniques that can be used for making a feature sequence more robust to variations in aspects such as local tempo, articulation, and note execution.\n",
"\n",
"To illustrate the effect of the various postprocessing techniques, we use in this notebook the beginning of Beethoven's Fifth Symphony as our running example. As for the recordings, we consider different real and synthesized performances of the piece including orchestral versions, piano versions, and a string quartet version. \n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"\n",
"
\n",
"
\n",
"
Performer
\n",
"
Instrumentation
\n",
"
Audio
\n",
"
\n",
" \n",
"
\n",
"
Bernstein
\n",
"
Orchestra
\n",
"
\n",
" \n",
"
\n",
"
\n",
" \n",
"
\n",
"
Karajan
\n",
"
Orchestra
\n",
"
\n",
" \n",
"
\n",
"
\n",
" \n",
"
\n",
"
MIDI-Orchestra
\n",
"
Orchestra
\n",
"
\n",
" \n",
"
\n",
"
\n",
" \n",
"
\n",
"
MIDI-Piano
\n",
"
Piano
\n",
"
\n",
" \n",
"
\n",
"
\n",
" \n",
"
\n",
"
Scherbakov
\n",
"
Piano
\n",
"
\n",
" \n",
"
\n",
"
\n",
" \n",
"
\n",
"
MIDI-Quartet
\n",
"
String quartet
\n",
"
\n",
" \n",
"
\n",
"
\n",
"
"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"The following figure shows chroma-based feature representations for some of these performances. The chromagrams are [normalized](../C3/C3S1_FeatureNormalization.html) with respect to the Euclidean norm. Even though different interpretations exhibit significant variations in articulation and instrumentation, the chromagrams show a similar progression of the chroma distribution over time. "
]
},
{
"cell_type": "code",
"execution_count": 1,
"metadata": {
"execution": {
"iopub.execute_input": "2024-02-15T08:50:59.161885Z",
"iopub.status.busy": "2024-02-15T08:50:59.161661Z",
"iopub.status.idle": "2024-02-15T08:51:05.156813Z",
"shell.execute_reply": "2024-02-15T08:51:05.156282Z"
}
},
"outputs": [
{
"data": {
"image/png": "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\n",
"text/plain": [
"
"
]
},
"metadata": {
"needs_background": "light"
},
"output_type": "display_data"
}
],
"source": [
"import os\n",
"import sys\n",
"import numpy as np\n",
"from matplotlib import pyplot as plt\n",
"import librosa\n",
"from scipy import signal\n",
"\n",
"sys.path.append('..')\n",
"import libfmp.b\n",
"import libfmp.c3\n",
"\n",
"%matplotlib inline\n",
"\n",
"fn_wav_dict = {}\n",
"fn_wav_dict['Bernstein'] = os.path.join('..', 'data', 'C3', 'FMP_C3S3_Beethoven_Fifth-MM1-21_Bernstein.wav')\n",
"fn_wav_dict['Karajan'] = os.path.join('..', 'data', 'C3', 'FMP_C3S3_Beethoven_Fifth-MM1-21_Karajan1946.wav')\n",
"#fn_wav_dict['MIDI-Orchestra'] = os.path.join('..', 'data', 'C3', 'FMP_C3S3_Beethoven_Fifth-MM1-21_Sibelius-Orchestra-Fast.wav')\n",
"fn_wav_dict['Scherbakov']= os.path.join('..', 'data', 'C3', 'FMP_C3S3_Beethoven_Fifth-MM1-21_Scherbakov.wav')\n",
"fn_wav_dict['MIDI-Piano']= os.path.join('..', 'data', 'C3', 'FMP_C3S3_Beethoven_Fifth-MM1-21_Sibelius-Piano.wav')\n",
"#fn_wav_dict['MIDI-Piano'] = os.path.join('..', 'data', 'C3', 'FMP_C3S3_Beethoven_Fifth-MM1-21_Midi-Piano.wav')\n",
"#fn_wav_dict['MIDI-Quartet']= os.path.join('..', 'data', 'C3', 'FMP_C3S3_Beethoven_Fifth-MM1-21_Sibelius-Quartett.wav')\n",
"\n",
"N, H = 2048, 1024\n",
"figsize = (8, 1.5)\n",
"yticks = [0, 4, 7, 11]\n",
"\n",
"C_dict = {}\n",
"for name in fn_wav_dict:\n",
" fn_wav = fn_wav_dict[name] \n",
" x, Fs = librosa.load(fn_wav)\n",
" C = librosa.feature.chroma_stft(y=x, sr=Fs, tuning=0, norm=None, hop_length=H, n_fft=N)\n",
" Fs_C = Fs / H\n",
" C = C / C.max()\n",
" threshold = 0.0001\n",
" C_dict[name] = libfmp.c3.normalize_feature_sequence(C, norm='2', threshold=threshold)\n",
" libfmp.b.plot_chromagram(C_dict[name], Fs_C, figsize=figsize, ylabel=name, xlabel='', chroma_yticks=yticks)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Temporal Smoothing and Downsampling\n",
"\n",
"For certain music retrieval applications, these chromagrams may be too detailed. In particular, it may be desirable to further increase the similarity between them. We now show how this can be achieved by **smoothing procedures** applied in a postprocessing step. The idea is to compute for each chroma dimension a kind of local average over time. More precisely, let $X=(x_1,x_2, ..., x_N)$ be a feature sequence with $x_n\\in\\mathbb{R}^K$ for $n\\in[1:N]$, and let $w$ be a [rectangular window](../C2/C2_STFT-Window.html) $w$ of length $L\\in\\mathbb{N}$. Then we compute for each $k\\in[1:K]$ a convolution between $w$ and the sequence $(x_1(k), x_2(k),\\ldots, x_N(k))$. Assuming a [centered view](../C2/C2_STFT-Conventions.html), we only keep the center part of length $N$ of the convolution. The result is a smoothed feature sequence of the same dimensions $K\\times N$.\n",
"\n",
"* In the following implementation, we use the function `scipy.signal.convolve` to compute a 2D convolution. Setting the dimension to $1\\times L$ for the **window** (also called **kernel**) results in a bandwise 1D convolution.\n",
"* Using the parameter `mode='same'` enforces the centered view.\n",
"* As for the window $w$, one may also use other [window types](../C2/C2_STFT-Window.html) such as a **Hann window**. \n",
"\n",
"Applying temporal smoothing using a rectangular or a Hann window can be regarded as bandwise **lowpass filtering**, which attenuates fast temporal fluctuations in the feature representation. Often, to increase the efficiency of subsequent processing and analysis steps, one decimates the smoothed representation by keeping only every $H^\\mathrm{th}$ feature, where $H\\in\\mathbb{N}$ is a suitable constant (typically much smaller than the window length $L$). This decimation, which is also referred to as **downsampling**, reduces the feature rate by a factor $H$. \n",
"\n",
"The following figure shows the effect of our smoothing procedure continuing our Beethoven example from above. After smoothing with a rectangular window of length $L=11$, we apply feature normalization using the Euclidean norm, and finally apply downsampling by $H=2$. "
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {
"execution": {
"iopub.execute_input": "2024-02-15T08:51:05.159495Z",
"iopub.status.busy": "2024-02-15T08:51:05.159251Z",
"iopub.status.idle": "2024-02-15T08:51:07.188074Z",
"shell.execute_reply": "2024-02-15T08:51:07.187498Z"
}
},
"outputs": [
{
"data": {
"image/png": "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\n",
"text/plain": [
"
"
]
},
"metadata": {
"needs_background": "light"
},
"output_type": "display_data"
}
],
"source": [
"def smooth_downsample_feature_sequence(X, Fs, filt_len=41, down_sampling=10, w_type='boxcar'):\n",
" \"\"\"Smoothes and downsamples a feature sequence. Smoothing is achieved by convolution with a filter kernel\n",
"\n",
" Notebook: C3/C3S1_FeatureSmoothing.ipynb\n",
"\n",
" Args:\n",
" X (np.ndarray): Feature sequence\n",
" Fs (scalar): Frame rate of ``X``\n",
" filt_len (int): Length of smoothing filter (Default value = 41)\n",
" down_sampling (int): Downsampling factor (Default value = 10)\n",
" w_type (str): Window type of smoothing filter (Default value = 'boxcar')\n",
"\n",
" Returns:\n",
" X_smooth (np.ndarray): Smoothed and downsampled feature sequence\n",
" Fs_feature (scalar): Frame rate of ``X_smooth``\n",
" \"\"\"\n",
" filt_kernel = np.expand_dims(signal.get_window(w_type, filt_len), axis=0)\n",
" X_smooth = signal.convolve(X, filt_kernel, mode='same') / filt_len\n",
" X_smooth = X_smooth[:, ::down_sampling]\n",
" Fs_feature = Fs / down_sampling\n",
" return X_smooth, Fs_feature\n",
"\n",
"filt_len = 11\n",
"down_sampling = 2\n",
"C_smooth_dict = {} \n",
"for name in fn_wav_dict: \n",
" C_smooth, Fs_C_smooth = smooth_downsample_feature_sequence(C_dict[name], Fs_C, \n",
" filt_len=filt_len, down_sampling=down_sampling)\n",
" C_smooth_dict[name] = libfmp.c3.normalize_feature_sequence(C_smooth, norm='2', threshold=threshold)\n",
" libfmp.b.plot_chromagram(C_smooth_dict[name], Fs_C_smooth, figsize=figsize, \n",
" ylabel=name, title='', xlabel='', chroma_yticks=yticks)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Smoothing via Median Filtering\n",
"\n",
"An alternative to applying a local averaging filter is **median filtering**, which also results in some smoothing while better preserving sharp transitions. The **median** of a finite list of numbers is the numerical value with the property that half the numbers fall below the value and half above it. The median can be computed by arranging all the numbers from lowest value to highest value and picking the middle one. If there is an even number of observations, then there is no single middle value; the median is then usually defined to be the mean of the two middle values. For example, the median of the list $(5,3,2,8,2)$ is $3$, while the median of the list $(5,3,2,8)$ is $4$. The idea of **median filtering** is to replace each entry of a sequence with the **median** of neighboring entries. The size of the neighborhood is determined by a parameter $L\\in\\mathbb{N}$, which is the **length** of the median filter applied. Given a sequence $X=(x_1,x_2, ..., x_N)$ of feature vectors $x_n\\in\\mathbb{R}^K$, we apply median filtering for each dimension for each $k\\in[1:K]$ (just as we did with average smoothing). \n",
"\n",
"* In the following implementation, we use the function `scipy.signal.medfilt2d` to compute a 2D median filtering. Using median filtering with kernel size $1\\times L$ results in a bandwise 1D median filtering.\n",
"* When using `scipy.signal.medfilt2d`, the median filter length $L$ needs to be odd.\n",
"* The output array of median filtering has the same dimensions as the input array. Boundary issues are handled by zero padding. \n",
"\n",
"In the following code cell, we give a toy example to illustrate the definition and properties of median filtering."
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {
"execution": {
"iopub.execute_input": "2024-02-15T08:51:07.190660Z",
"iopub.status.busy": "2024-02-15T08:51:07.190469Z",
"iopub.status.idle": "2024-02-15T08:51:07.195073Z",
"shell.execute_reply": "2024-02-15T08:51:07.194512Z"
}
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Input array X of dimension (K,N) with K=3 and N=5\n",
"[[1. 2. 3. 4. 5.]\n",
" [5. 6. 7. 8. 9.]\n",
" [5. 3. 2. 8. 2.]]\n",
"Output array after median filtering with L=3\n",
"[[1. 2. 3. 4. 4.]\n",
" [5. 6. 7. 8. 8.]\n",
" [3. 3. 3. 2. 2.]]\n"
]
}
],
"source": [
"X = np.array([[1, 2, 3, 4, 5], [5, 6, 7, 8, 9], [5, 3, 2, 8, 2]], dtype='float')\n",
"L = 3\n",
"filt_len = [1, L]\n",
"X_smooth = signal.medfilt2d(X, filt_len)\n",
"print('Input array X of dimension (K,N) with K=3 and N=5')\n",
"print(X)\n",
"print('Output array after median filtering with L=3')\n",
"print(X_smooth)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Continuing our Beethoven example, the following figure compares the original chromagram with its smoothed versions—once applying average filtering and once median filtering. In the original chromagram, one can observe sharp edges that are the result of note transition (e.g. look at the two fate motives at the beginning with transitions $\\mathrm{G}$, $\\mathrm{E}^\\flat$, $\\mathrm{F}$, $\\mathrm{D}$). While average smoothing introduces a smearing between these transitions, median smoothing tends to better preserve the sharp edges. "
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {
"execution": {
"iopub.execute_input": "2024-02-15T08:51:07.197500Z",
"iopub.status.busy": "2024-02-15T08:51:07.197317Z",
"iopub.status.idle": "2024-02-15T08:51:08.015919Z",
"shell.execute_reply": "2024-02-15T08:51:08.015283Z"
}
},
"outputs": [
{
"data": {
"image/png": "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\n",
"text/plain": [
"
"
]
},
"metadata": {
"needs_background": "light"
},
"output_type": "display_data"
}
],
"source": [
"def median_downsample_feature_sequence(X, Fs, filt_len=41, down_sampling=10):\n",
" \"\"\"Smoothes and downsamples a feature sequence. Smoothing is achieved by median filtering\n",
"\n",
" Notebook: C3/C3S1_FeatureSmoothing.ipynb\n",
"\n",
" Args:\n",
" X (np.ndarray): Feature sequence\n",
" Fs (scalar): Frame rate of ``X``\n",
" filt_len (int): Length of smoothing filter (Default value = 41)\n",
" down_sampling (int): Downsampling factor (Default value = 10)\n",
"\n",
" Returns:\n",
" X_smooth (np.ndarray): Smoothed and downsampled feature sequence\n",
" Fs_feature (scalar): Frame rate of ``X_smooth``\n",
" \"\"\"\n",
" assert filt_len % 2 == 1 # L needs to be odd\n",
" filt_len = [1, filt_len]\n",
" X_smooth = signal.medfilt2d(X, filt_len)\n",
" X_smooth = X_smooth[:, ::down_sampling]\n",
" Fs_feature = Fs / down_sampling\n",
" return X_smooth, Fs_feature\n",
"\n",
"filt_len = 11\n",
"down_sampling = 2\n",
"C_median_dict = {} \n",
"for name in fn_wav_dict: \n",
" C_median, Fs_C_smooth = median_downsample_feature_sequence(C_dict[name], Fs_C,\n",
" filt_len=filt_len, down_sampling=down_sampling)\n",
" C_median_dict[name] = libfmp.c3.normalize_feature_sequence(C_median, norm='2', threshold=threshold)\n",
"\n",
"figsize=(8, 1.7)\n",
"\n",
"name = 'Karajan'\n",
"libfmp.b.plot_chromagram(C_dict[name], Fs_C_smooth, figsize=figsize, ylabel = name,\n",
" title='Original chromagram', xlabel='', chroma_yticks=yticks)\n",
"libfmp.b.plot_chromagram(C_smooth_dict[name], Fs_C_smooth, figsize=figsize, ylabel = name,\n",
" title='Smoothed chromagram using average filtering', xlabel='', chroma_yticks=yticks) \n",
"libfmp.b.plot_chromagram(C_median_dict[name], Fs_C_smooth, figsize=figsize, ylabel = name, \n",
" title='Smoothed chromagram using median filtering', xlabel='', chroma_yticks=yticks) \n",
"\n",
"C_diff = C_smooth_dict[name] - C_median_dict[name]\n",
"m = np.max(np.abs(C_diff))\n",
"libfmp.b.plot_chromagram(C_diff, Fs_C_smooth, cmap='seismic', clim=[-m, m], figsize=figsize,\n",
" title='Difference between average- and median-filtered chromagram', xlabel='',\n",
" ylabel=name, chroma_yticks=yticks);"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Further Notes\n",
"\n",
"In summary, there are many ways to enhance and modify a feature representation (e.g., a chromagram) by applying techniques such as [logarithmic compression](../C3/C3S1_LogCompression.html), [feature normalization](../C3/C3S1_FeatureNormalization.html), smoothing, and downsampling. The described techniques provide flexible and computationally inexpensive tools for adjusting the feature specificity and resolution without repeating the cost-intensive spectral audio decomposition. Here are some further comments and links:\n",
"\n",
"* An alternative to temporal smoothing based on convolution with a window of fixed length is the usage of a [beat-synchronous feature representation](../C6/C6S3_AdaptiveWindowing.html). In this case, a windowed section is determined by two consecutive beat positions, which results in one feature vector per beat. This technique will be discuss in the [FMP notebook on adaptive windowing](../C6/C6S3_AdaptiveWindowing.html).\n",
"\n",
"* Given a short query audio clip and a music database, the task of [audio matching](../C7/C7S2_AudioMatching.html) aims at retrieving from the database all audio excerpts that are musically related to the query. Within this application scenario, enhanced and smoothed chroma-based audio features are studied in the [FMP notebook on feature design](../C7/C7S2_CENS.html). "
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"
![\"FMP\"](\"../data/FMP_Teaser_Cover.png\")
\n",
"![\"AudioLabs\"](\"../data/Logo_AudioLabs_Long.png\"){kind=logo}
\n",
"![\"C3\"](\"../data/C3_nav.png\")
\n",
"![\"C3\"](\"../data/C3/FMP_C3S3_Beethoven_Fifth-MM1-21_Sibelius-Piano.png\")
\n",
"\n",
"\n",
"\n",
"![\"C0\"](\"../data/C0_nav.png\")
![\"C1\"](\"../data/C1_nav.png\")
![\"C2\"](\"../data/C2_nav.png\")
![\"C4\"](\"../data/C4_nav.png\")
![\"C5\"](\"../data/C5_nav.png\")
![\"C6\"](\"../data/C6_nav.png\")
![\"C7\"](\"../data/C7_nav.png\")
![\"C8\"](\"../data/C8_nav.png\")