[Python] Flattening signal frames withou overlap after applying VAD Filter using Librosa

Hello I've built a very basic VAD filter using Librosa.

# Size of our frame is 552 samples
frame_len = 552
# Size of the overlap is 221 samples
frame_hop = 221

# Slices data array into overlapping frames without previously defined variables
frames = librosa.util.frame(y,
frame_length=frame_len,
hop_length=frame_hop)

# Computing average loudness in frame.
# summing the square of our frame and dividing by frame length
# Squaring the Signal positive and negative don't cancel each other out.
energy = np.sum(frames**2, axis=0) / frame_len

# Convert to decibels and add a tiny (1e-10) constant to avoid log(0) errors
# Since the above calculates power we multiply by 10 (not 20 as would be the case for Amplitude) [1]
# To obtain the average decibel power of our frame
energy_db = 10 * np.log10(energy + 1e-10)

# Comparing the loudest frame energy_db.max() to other frames
# anything not louder than energy_db.max() - 40db, is cut off
# We return a Boolean Array
vad_mask = energy_db > (energy_db.max() - 40)



The issue I'm running into is that when I'm just naively filtering and flattening the signal, I keep the overlap. (Effectively this sounds like I'm lowering the frequency)

# We select only the vad_mask frames
speech_frames = frames[:, vad_mask]
# Transpose and flatten the frames to yield one continious signal
y_vad = speech_frames.T.reshape(-1)


I could simply remove the preceding and trailing 221 samples from each frame, but I'm wondering if there is a librosa function which will help me do that.

Continue reading...