| """
|
| ===============================================================================
|
| preprocessing/silence_removal.py — Energy-Based Silence Trimming
|
| ===============================================================================
|
|
|
| """
|
|
|
| import numpy as np
|
| import librosa
|
| from config import SILENCE_TOP_DB
|
|
|
|
|
| def remove_silence(audio, sr):
|
| """
|
| Remove leading and trailing silence from the audio signal.
|
|
|
| Uses energy-based detection to find where the "real" sound starts and ends,
|
| then trims everything outside that window.
|
|
|
| Parameters
|
| ----------
|
| audio : np.ndarray
|
| 1D array of audio samples.
|
| sr : int
|
| Sampling rate of the audio.
|
|
|
| Returns
|
| -------
|
| np.ndarray
|
| Trimmed audio signal. If the entire signal is below the threshold,
|
| returns the original audio unchanged (safety fallback).
|
|
|
| Notes
|
| -----
|
| - top_db is imported from config.py (default: 20 dB).
|
| - A small margin (frame_length=2048, hop_length=512) is used for
|
| energy estimation. These are independent of the mel spectrogram
|
| parameters — they only control how finely we detect silence boundaries.
|
|
|
| """
|
|
|
|
|
| trimmed_audio, index = librosa.effects.trim(
|
| audio,
|
| top_db=SILENCE_TOP_DB,
|
| frame_length = 2048,
|
| hop_length = 512
|
| )
|
| if len(trimmed_audio) < 1024:
|
| return audio
|
|
|
| return trimmed_audio
|
|
|
