def compute_melmat(self):
return mel.compute_melmat(
num_mel_bands=self._config['samples'],
freq_min=self._config['min_frequency'],
freq_max=self._config['max_frequency'],
num_fft_bands=int(self._config['nfft'] // 2) + 1,
sample_rate=self._config['mic_rate'])
def _initialize_melbank(self):
"""Initialize all the melbank related variables"""
# Few difference coefficient types for experimentation
if self._config["coeffs_type"] == "triangle":
melbank_mel = np.linspace(
aubio.hztomel(self._config["min_frequency"]),
aubio.hztomel(self._config["max_frequency"]),
self._config["samples"] + 2,
)
self.melbank_frequencies = np.array(
[aubio.meltohz(mel) for mel in melbank_mel]
).astype(np.float32)
self.filterbank = aubio.filterbank(
self._config["samples"], self._config["fft_size"]
)
self.filterbank.set_triangle_bands(
self.melbank_frequencies, self._config["mic_rate"]
)
self.melbank_frequencies = self.melbank_frequencies[1:-1]
if self._config["coeffs_type"] == "bark":
melbank_bark = np.linspace(
6.0 * np.arcsinh(self._config["min_frequency"] / 600.0),
6.0 * np.arcsinh(self._config["max_frequency"] / 600.0),
self._config["samples"] + 2,
)
self.melbank_frequencies = (
600.0 * np.sinh(melbank_bark / 6.0)
).astype(np.float32)
self.filterbank = aubio.filterbank(
self._config["samples"], self._config["fft_size"]
)
self.filterbank.set_triangle_bands(
self.melbank_frequencies, self._config["mic_rate"]
)
self.melbank_frequencies = self.melbank_frequencies[1:-1]
# Slaney coefficients will always produce 40 samples spanning 133Hz to
# 6000Hz
if self._config["coeffs_type"] == "slaney":
self.filterbank = aubio.filterbank(40, self._config["fft_size"])
self.filterbank.set_mel_coeffs_slaney(self._config["mic_rate"])
# Sanley frequencies are linear-log spaced where 133Hz to 1000Hz is linear
# spaced and 1000Hz to 6000Hz is log spaced. It also produced a hardcoded
# 40 samples.
lowestFrequency = 133.3
linearSpacing = 66.6666666
logSpacing = 1.0711703
linearFilters = 13
logFilters = 27
linearSpacedFreqs = (
lowestFrequency + np.arange(0, linearFilters) * linearSpacing
)
logSpacedFreqs = linearSpacedFreqs[-1] * np.power(
logSpacing, np.arange(1, logFilters + 1)
)
self._config["samples"] = 40
self.melbank_frequencies = np.hstack(
(linearSpacedFreqs, logSpacedFreqs)
).astype(np.float32)
# Standard mel coefficients
if self._config["coeffs_type"] == "mel":
self.filterbank = aubio.filterbank(
self._config["samples"], self._config["fft_size"]
)
self.filterbank.set_mel_coeffs(
self._config["mic_rate"],
self._config["min_frequency"],
self._config["max_frequency"],
)
# Frequencies wil be linearly spaced in the mel scale
melbank_mel = np.linspace(
aubio.hztomel(self._config["min_frequency"]),
aubio.hztomel(self._config["max_frequency"]),
self._config["samples"],
)
self.melbank_frequencies = np.array(
[aubio.meltohz(mel) for mel in melbank_mel]
)
# HTK mel coefficients
if self._config["coeffs_type"] == "htk":
self.filterbank = aubio.filterbank(
self._config["samples"], self._config["fft_size"]
)
self.filterbank.set_mel_coeffs_htk(
self._config["mic_rate"],
self._config["min_frequency"],
self._config["max_frequency"],
)
# Frequencies wil be linearly spaced in the mel scale
melbank_mel = np.linspace(
aubio.hztomel(self._config["min_frequency"]),
aubio.hztomel(self._config["max_frequency"]),
self._config["samples"],
)
self.melbank_frequencies = np.array(
[aubio.meltohz(mel) for mel in melbank_mel]
)
# Coefficients based on Scott's audio reactive led project
if self._config["coeffs_type"] == "scott":
(melmat, center_frequencies_hz, freqs,) = mel.compute_melmat(
num_mel_bands=self._config["samples"],
freq_min=self._config["min_frequency"],
freq_max=self._config["max_frequency"],
num_fft_bands=int(self._config["fft_size"] // 2) + 1,
sample_rate=self._config["mic_rate"],
)
self.filterbank = aubio.filterbank(
self._config["samples"], self._config["fft_size"]
)
self.filterbank.set_coeffs(melmat.astype(np.float32))
self.melbank_frequencies = center_frequencies_hz
# "Mel"-spacing based on Scott's audio reactive led project. This
# should in theory be the same as the above, but there seems to be
# slight differences. Leaving both for science!
if self._config["coeffs_type"] == "scott_mel":
def hertz_to_scott(freq):
return 3340.0 * log(1 + (freq / 250.0), 9)
def scott_to_hertz(scott):
return 250.0 * (9 ** (scott / 3340.0)) - 250.0
melbank_scott = np.linspace(
hertz_to_scott(self._config["min_frequency"]),
hertz_to_scott(self._config["max_frequency"]),
self._config["samples"] + 2,
)
self.melbank_frequencies = np.array(
[scott_to_hertz(scott) for scott in melbank_scott]
).astype(np.float32)
self.filterbank = aubio.filterbank(
self._config["samples"], self._config["fft_size"]
)
self.filterbank.set_triangle_bands(
self.melbank_frequencies, self._config["mic_rate"]
)
self.melbank_frequencies = self.melbank_frequencies[1:-1]
# Modified scott_mel, spreads out the low range and compresses the
# highs
if self._config["coeffs_type"] == "matt_mel":
def hertz_to_matt(freq):
return 3700.0 * log(1 + (freq / 200.0), 13)
def matt_to_hertz(matt):
return 200.0 * (10 ** (matt / 3700.0)) - 200.0
melbank_matt = np.linspace(
hertz_to_matt(self._config["min_frequency"]),
hertz_to_matt(self._config["max_frequency"]),
self._config["samples"] + 2,
)
self.melbank_frequencies = np.array(
[matt_to_hertz(matt) for matt in melbank_matt]
).astype(np.float32)
self.filterbank = aubio.filterbank(
self._config["samples"], self._config["fft_size"]
)
self.filterbank.set_triangle_bands(
self.melbank_frequencies, self._config["mic_rate"]
)
self.melbank_frequencies = self.melbank_frequencies[1:-1]
if self._config["coeffs_type"] == "fixed":
ranges = FREQUENCY_RANGES.values()
upper_edges_hz = np.zeros(len(ranges))
lower_edges_hz = np.zeros(len(ranges))
for idx, value in enumerate(ranges):
lower_edges_hz[idx] = value.min
upper_edges_hz[idx] = value.max
(
melmat,
center_frequencies_hz,
freqs,
) = mel.compute_melmat_from_range(
lower_edges_hz=lower_edges_hz,
upper_edges_hz=upper_edges_hz,
num_fft_bands=int(self._config["fft_size"] // 2) + 1,
sample_rate=self._config["mic_rate"],
)
self._config["samples"] = len(center_frequencies_hz)
self.filterbank = aubio.filterbank(
self._config["samples"], self._config["fft_size"]
)
self.filterbank.set_coeffs(melmat.astype(np.float32))
self.melbank_frequencies = center_frequencies_hz
if self._config["coeffs_type"] == "fixed_simple":
ranges = FREQUENCY_RANGES_SIMPLE.values()
upper_edges_hz = np.zeros(len(ranges))
lower_edges_hz = np.zeros(len(ranges))
for idx, value in enumerate(ranges):
lower_edges_hz[idx] = value.min
upper_edges_hz[idx] = value.max
(
melmat,
center_frequencies_hz,
freqs,
) = mel.compute_melmat_from_range(
lower_edges_hz=lower_edges_hz,
upper_edges_hz=upper_edges_hz,
num_fft_bands=int(self._config["fft_size"] // 2) + 1,
sample_rate=self._config["mic_rate"],
)
self._config["samples"] = len(center_frequencies_hz)
self.filterbank = aubio.filterbank(
self._config["samples"], self._config["fft_size"]
)
self.filterbank.set_coeffs(melmat.astype(np.float32))
self.melbank_frequencies = center_frequencies_hz
self.melbank_frequencies = self.melbank_frequencies.astype(int)
# Normalize the filterbank triangles to a consistent height, the
# default coeffs (for types other than legacy) will be normalized
# by the triangles area which results in an uneven melbank
if (
self._config["coeffs_type"] != "scott"
and self._config["coeffs_type"] == "scott_mel"
):
coeffs = self.filterbank.get_coeffs()
coeffs /= np.max(coeffs, axis=-1)[:, None]
self.filterbank.set_coeffs(coeffs)
# Find the indexes for each of the frequency ranges
self.lows_index = self.mids_index = self.highs_index = 1
for i in range(0, len(self.melbank_frequencies)):
if (
self.melbank_frequencies[i]
< FREQUENCY_RANGES_SIMPLE["Low (1-250Hz)"].max
):
self.lows_index = i + 1
elif (
self.melbank_frequencies[i]
< FREQUENCY_RANGES_SIMPLE["Mid (250Hz-4kHz)"].max
):
self.mids_index = i + 1
elif (
self.melbank_frequencies[i]
< FREQUENCY_RANGES_SIMPLE["High (4kHz-24kHz)"].max
):
self.highs_index = i + 1
# Build up some of the common filters
self.mel_gain = ExpFilter(
np.tile(1e-1, self._config["samples"]),
alpha_decay=0.01,
alpha_rise=0.99,
)
self.mel_smoothing = ExpFilter(
np.tile(1e-1, self._config["samples"]),
alpha_decay=0.2,
alpha_rise=0.99,
)
self.common_filter = ExpFilter(alpha_decay=0.99, alpha_rise=0.01)
def _initialize_melbank(self):
"""Initialize all the melbank related variables"""
# Few difference coefficient types for experimentation
if self._config['coeffs_type'] == 'triangle':
melbank_mel = np.linspace(
aubio.hztomel(self._config['min_frequency']),
aubio.hztomel(self._config['max_frequency']),
self._config['samples'] + 2)
self.melbank_frequencies = np.array(
[aubio.meltohz(mel) for mel in melbank_mel]).astype(np.float32)
self.filterbank = aubio.filterbank(self._config['samples'],
self._config['fft_size'])
self.filterbank.set_triangle_bands(self.melbank_frequencies,
self._config['mic_rate'])
self.melbank_frequencies = self.melbank_frequencies[1:-1]
if self._config['coeffs_type'] == 'bark':
melbank_bark = np.linspace(
6.0 * np.arcsinh(self._config['min_frequency'] / 600.0),
6.0 * np.arcsinh(self._config['max_frequency'] / 600.0),
self._config['samples'] + 2)
self.melbank_frequencies = (600.0 *
np.sinh(melbank_bark / 6.0)).astype(
np.float32)
self.filterbank = aubio.filterbank(self._config['samples'],
self._config['fft_size'])
self.filterbank.set_triangle_bands(self.melbank_frequencies,
self._config['mic_rate'])
self.melbank_frequencies = self.melbank_frequencies[1:-1]
# Slaney coefficients will always produce 40 samples spanning 133Hz to 6000Hz
if self._config['coeffs_type'] == 'slaney':
self.filterbank = aubio.filterbank(40, self._config['fft_size'])
self.filterbank.set_mel_coeffs_slaney(self._config['mic_rate'])
# Sanley frequencies are linear-log spaced where 133Hz to 1000Hz is linear
# spaced and 1000Hz to 6000Hz is log spaced. It also produced a hardcoded
# 40 samples.
lowestFrequency = 133.3
linearSpacing = 66.6666666
logSpacing = 1.0711703
linearFilters = 13
logFilters = 27
linearSpacedFreqs = lowestFrequency + np.arange(
0, linearFilters) * linearSpacing
logSpacedFreqs = linearSpacedFreqs[-1] * np.power(
logSpacing, np.arange(1, logFilters + 1))
self._config['samples'] = 40
self.melbank_frequencies = np.hstack(
(linearSpacedFreqs, logSpacedFreqs)).astype(np.float32)
# Standard mel coefficients
if self._config['coeffs_type'] == 'mel':
self.filterbank = aubio.filterbank(self._config['samples'],
self._config['fft_size'])
self.filterbank.set_mel_coeffs(self._config['mic_rate'],
self._config['min_frequency'],
self._config['max_frequency'])
# Frequencies wil be linearly spaced in the mel scale
melbank_mel = np.linspace(
aubio.hztomel(self._config['min_frequency']),
aubio.hztomel(self._config['max_frequency']),
self._config['samples'])
self.melbank_frequencies = np.array(
[aubio.meltohz(mel) for mel in melbank_mel])
# HTK mel coefficients
if self._config['coeffs_type'] == 'htk':
self.filterbank = aubio.filterbank(self._config['samples'],
self._config['fft_size'])
self.filterbank.set_mel_coeffs_htk(self._config['mic_rate'],
self._config['min_frequency'],
self._config['max_frequency'])
# Frequencies wil be linearly spaced in the mel scale
melbank_mel = np.linspace(
aubio.hztomel(self._config['min_frequency']),
aubio.hztomel(self._config['max_frequency']),
self._config['samples'])
self.melbank_frequencies = np.array(
[aubio.meltohz(mel) for mel in melbank_mel])
# Coefficients based on Scott's audio reactive led project
if self._config['coeffs_type'] == 'scott':
(melmat, center_frequencies_hz, freqs) = mel.compute_melmat(
num_mel_bands=self._config['samples'],
freq_min=self._config['min_frequency'],
freq_max=self._config['max_frequency'],
num_fft_bands=int(self._config['fft_size'] // 2) + 1,
sample_rate=self._config['mic_rate'])
self.filterbank = aubio.filterbank(self._config['samples'],
self._config['fft_size'])
self.filterbank.set_coeffs(melmat.astype(np.float32))
self.melbank_frequencies = center_frequencies_hz
# "Mel"-spacing based on Scott's audio reactive led project. This
# should in theory be the same as the above, but there seems to be
# slight differences. Leaving both for science!
if self._config['coeffs_type'] == 'scott_mel':
def hertz_to_scott(freq):
return 3340.0 * log(1 + (freq / 250.0), 9)
def scott_to_hertz(scott):
return 250.0 * (9**(scott / 3340.0)) - 250.0
melbank_scott = np.linspace(
hertz_to_scott(self._config['min_frequency']),
hertz_to_scott(self._config['max_frequency']),
self._config['samples'] + 2)
self.melbank_frequencies = np.array([
scott_to_hertz(scott) for scott in melbank_scott
]).astype(np.float32)
self.filterbank = aubio.filterbank(self._config['samples'],
self._config['fft_size'])
self.filterbank.set_triangle_bands(self.melbank_frequencies,
self._config['mic_rate'])
self.melbank_frequencies = self.melbank_frequencies[1:-1]
self.melbank_frequencies = self.melbank_frequencies.astype(int)
# Normalize the filterbank triangles to a consistent height, the
# default coeffs (for types other than legacy) will be normalized
# by the triangles area which results in an uneven melbank
if self._config['coeffs_type'] != 'scott' and self._config[
'coeffs_type'] == 'scott_mel':
coeffs = self.filterbank.get_coeffs()
coeffs /= np.max(coeffs, axis=-1)[:, None]
self.filterbank.set_coeffs(coeffs)
# Find the indexes for each of the frequency ranges
for i in range(0, len(self.melbank_frequencies) - 1):
if self.melbank_frequencies[i] < FREQUENCY_RANGES_SIMPLE['low'].max:
self.lows_index = i
elif self.melbank_frequencies[i] < FREQUENCY_RANGES_SIMPLE[
'mid'].max:
self.mids_index = i
elif self.melbank_frequencies[i] < FREQUENCY_RANGES_SIMPLE[
'high'].max:
self.highs_index = i
# Build up some of the common filters
self.mel_gain = ExpFilter(np.tile(1e-1, self._config['samples']),
alpha_decay=0.01,
alpha_rise=0.99)
self.mel_smoothing = ExpFilter(np.tile(1e-1, self._config['samples']),
alpha_decay=0.2,
alpha_rise=0.99)
self.common_filter = ExpFilter(alpha_decay=0.99, alpha_rise=0.01)