def get_sines_per_frame(audio, sr=44100, onlyfrecuencies=False, nsines=20):
"""
Perform framewise sinusoidal model in an audio
:param audio: Audio either mono or stereo. Will be downsampled to mono
:param sr: Samplerate used for the audio
:return: Nx2x100. N is the number of resulting frames. 2x100 are the frequencies and magnitudes respectively.
"""
if audio.ndim > 1:
audio = std.MonoMixer()(audio, audio.shape[1])
len_arrays = 0
for i, _ in enumerate(
std.FrameGenerator(audio, frameSize=4096, hopSize=2048)):
len_arrays = i
fft_algo = std.FFT()
sine_anal = std.SineModelAnal(maxnSines=nsines,
orderBy='frequency',
minFrequency=1)
sines = np.zeros([len_arrays + 1, 2, nsines], dtype=np.float32) + eps
for i, frame in enumerate(
std.FrameGenerator(audio, frameSize=4096, hopSize=2048)):
fft = fft_algo(frame)
freqs, mags, _ = sine_anal(fft)
sorting_indexes = np.argsort(freqs)
freqs = freqs[sorting_indexes]
mags = mags[sorting_indexes]
sines[i, :] = [freqs, mags]
if onlyfrecuencies:
return sines[:, 0, :]
else:
return sines[:, 0, :], sines[:, 1, :]
def get_hpeaks_per_frame(audio, sr=44100, onlyfrecuencies=False, nsines=20):
"""
Get Harmonic peaks in an audio
:param audio: Audio either mono or stereo. Will be downsampled to mono
:param sr: Samplerate used for the audio
:return: Nx2x100. N is the number of resulting frames. 2x100 are the frequencies and magnitudes respectively.
"""
if audio.ndim > 1:
audio = std.MonoMixer()(audio, audio.shape[1])
fft_algo = std.FFT()
pyin = std.PitchYin()
hpeaks = std.HarmonicPeaks()
sine_anal = std.SineModelAnal(maxnSines=nsines,
orderBy='frequency',
minFrequency=1)
sines = []
for i, frame in enumerate(
std.FrameGenerator(audio, frameSize=4096, hopSize=2048)):
pitch, _ = pyin(frame)
fft = fft_algo(frame)
freqs, mags, _ = sine_anal(fft)
sorting_indexes = np.argsort(freqs)
freqs = freqs[sorting_indexes]
mags = mags[sorting_indexes]
non_zero_freqs = np.where(freqs != 0)
freqs = freqs[non_zero_freqs]
mags = mags[non_zero_freqs]
freqs, mags = hpeaks(freqs, mags, pitch)
sines.append([freqs, mags])
sines = np.array(sines)
if onlyfrecuencies:
return sines[:, 0, :]
else:
return sines[:, 0, :], sines[:, 1, :]
def mix_to_mono(audio):
"""
Mix an audio file down to mono
:param audio: (np.ndarray) audio samples
:return: (nd.ndarray) mono audio samples
"""
mono_mix = es.MonoMixer()
samples = mono_mix(audio, audio.shape[1])
return samples
def mix(audio1, audio2, sr):
"""
Function to mix audios with a normalised loudness
:param audio1: Audio vector to normalize
:param audio2: Audio vector to normalize
:param sr: Sample rate of the final mix
:return: Audio vector of the normalised mix
"""
if audio1.ndim > 1:
audio1 = std.MonoMixer()(audio1, audio1.shape[1])
if audio2.ndim > 1:
audio2 = std.MonoMixer()(audio2, audio2.shape[1])
std.MonoWriter(filename='temporal1.wav', sampleRate=sr)(audio1)
std.MonoWriter(filename='temporal2.wav', sampleRate=sr)(audio2)
stream1 = (ffmpeg.input('temporal1.wav').filter('loudnorm'))
stream2 = (ffmpeg.input('temporal2.wav').filter('loudnorm'))
merged_audio = ffmpeg.filter([stream1, stream2], 'amix')
ffmpeg.output(merged_audio, 'temporal_o.wav').overwrite_output().run()
audio_numpy = std.MonoLoader(filename='temporal_o.wav')()
return audio_numpy
# Compute beat positions and BPM
rhythm_extractor = esst.RhythmExtractor2013(method="multifeature")
bpm, beats, beats_confidence, _, beats_intervals = rhythm_extractor(
audio)
# print("BPM:", bpm)
# print("Beat positions (sec.):", beats)
print("Beat estimation confidence:", beats_confidence)
for i in range(len(beats)):
trim = esst.Trimmer(startTime=[*map(lambda x: x - 0.01, beats)][i],
endTime=[*map(lambda x: x + 0.15, beats)][i]) \
(audio)
if len(mixed_sample):
trim = np.resize(trim, mixed_sample.size)
stereo_mix = esst.StereoMuxer()(mixed_sample, trim)
esst.MonoMixer()(stereo_mix, 2)
else:
mixed_sample = trim
output_array = np.concatenate([output_array, mixed_sample])
esst.MonoWriter(filename=f"../samples/mix_beat{str(i)}.mp3")(output_array)
# Mark beat positions on the audio and write it to a file
# Let's use beeps instead of white noise to mark them, as it's more distinctive
# marker = AudioOnsetsMarker(onsets=beats, type='beep')
# marked_audio = marker(audio)
# MonoWriter(filename='../samples/dubstep_beats.flac')(marked_audio)