File size: 2,556 Bytes
1919bbe | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 | """
===============================================================================
preprocessing/__init__.py — Audio Preprocessing Pipeline
===============================================================================
This package contains all audio-level preprocessing steps that happen BEFORE
feature extraction (mel spectrograms). The pipeline order is:
Raw WAV → Resampling (JSON) → Normalization (JSON)
→ Noise Reduction (EL sir) → Silence Removal (EL sir)
→ Clean numpy array ready for feature extraction (sala7)
OWNERS: EL sir (noise_reduction, silence_removal)
JSON (normalization, resampling)
===============================================================================
"""
from preprocessing.resampling import resample_audio
from preprocessing.normalization import normalize_volume
from preprocessing.noise_reduction import reduce_noise
from preprocessing.silence_removal import remove_silence
def preprocess_audio(file_path):
"""
Full preprocessing pipeline: load → resample → normalize → denoise → trim.
This is the main entry point that JSON's Dataset class and infer.py will call.
It chains all four preprocessing steps in the correct order and returns
a clean numpy array ready for mel spectrogram extraction.
Parameters
----------
file_path : str
Absolute path to the .wav file.
Returns
-------
audio : np.ndarray
1D numpy array of the preprocessed audio waveform.
sr : int
Sampling rate (always TARGET_SR after resampling).
Pipeline Order & Rationale
--------------------------
1. Resample FIRST — all downstream steps assume a fixed sample rate.
If we denoise before resampling, the noise profile frequency bins
won't match after resampling.
2. Normalize SECOND — bring all recordings to the same loudness level
before noise estimation, so the noise threshold is consistent.
3. Denoise THIRD — remove background hum/motor noise.
4. Trim LAST — now that noise is gone, energy-based silence detection
can accurately find where the real machine sound starts/ends.
"""
# Step 1: Load and resample to TARGET_SR
audio, sr = resample_audio(file_path)
# Step 2: Volume normalization (peak or RMS)
audio = normalize_volume(audio)
# Step 3: Noise reduction
audio = reduce_noise(audio, sr)
# Step 4: Silence removal (trim leading/trailing silence)
audio = remove_silence(audio, sr)
return audio, sr
|