Gurkengewuerz/audio-reactive-led-strip
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Add zero padded FFT calculation, could help #25

Browse Source 
Zero pads the audio time samples until the length is equal to the next largest power of two. This improves the algorithmic complexity of the FFT calculations.

Initialized the Hamming window when the module is loaded instead of every loop.

Replaced a call to numpy.roll() with direct array index manipulation.
This commit is contained in:
Scott Lawson 2017-01-29 00:41:13 -08:00 committed by GitHub
parent 875943d6a1
commit fba8175e99
1 changed files with 8 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
python/visualization.py
View File

@ -167,7 +167,7 @@ mel_smoothing = dsp.ExpFilter(np.tile(1e-1, config.N_FFT_BINS),
alpha_decay=0.5, alpha_rise=0.99) alpha_decay=0.5, alpha_rise=0.99)
volume = dsp.ExpFilter(config.MIN_VOLUME_THRESHOLD, volume = dsp.ExpFilter(config.MIN_VOLUME_THRESHOLD,
alpha_decay=0.02, alpha_rise=0.02) alpha_decay=0.02, alpha_rise=0.02)
fft_window = np.hamming(int(config.MIC_RATE / config.FPS) * config.N_ROLLING_HISTORY)
# Keeps track of the number of buffer overflows # Keeps track of the number of buffer overflows
# Lots of buffer overflows could mean that FPS is set too high # Lots of buffer overflows could mean that FPS is set too high
buffer_overflows = 1 buffer_overflows = 1
@ -186,7 +186,7 @@ def microphone_update(stream):
# Normalize samples between 0 and 1 # Normalize samples between 0 and 1
y = y / 2.0**15 y = y / 2.0**15
# Construct a rolling window of audio samples # Construct a rolling window of audio samples
y_roll = np.roll(y_roll, -1, axis=0) y_roll[:-1] = y_roll[1:]
y_roll[-1, :] = np.copy(y) y_roll[-1, :] = np.copy(y)
y_data = np.concatenate(y_roll, axis=0) y_data = np.concatenate(y_roll, axis=0)
volume.update(np.nanmean(y_data ** 2)) volume.update(np.nanmean(y_data ** 2))
@ -197,10 +197,12 @@ def microphone_update(stream):
led.update() led.update()
else: else:
# Transform audio input into the frequency domain # Transform audio input into the frequency domain
XS, YS = dsp.fft(y_data, window=np.hamming) N = len(y_data)
# Remove half of the FFT data because of symmetry N_zeros = 2**int(np.ceil(np.log2(N))) - N
YS = YS[:len(YS) // 2] # Pad with zeros until the next power of two
XS = XS[:len(XS) // 2] y_data *= fft_window
y_padded = np.pad(y_data, (0, N_zeros), mode='constant')
YS = np.abs(np.fft.rfft(y_padded)[:N // 2])
# Construct a Mel filterbank from the FFT data # Construct a Mel filterbank from the FFT data
mel = np.atleast_2d(np.abs(YS)).T * dsp.mel_y.T mel = np.atleast_2d(np.abs(YS)).T * dsp.mel_y.T
# Scale data to values more suitable for visualization # Scale data to values more suitable for visualization