audiodiff.py

import numpy as np
import soundfile as sf
import librosa

file_a = "a.wav"
file_b = "b.wav"

# config
target_sr = 16000
chunk_ms = 50
threshold = 0.02
min_gap_ms = 100

# load & normalize
a, sr_a = sf.read(file_a, dtype='float32')
b, sr_b = sf.read(file_b, dtype='float32')
if a.ndim > 1: a = a.mean(axis=1)
if b.ndim > 1: b = b.mean(axis=1)
if sr_a != target_sr: a = librosa.resample(a, orig_sr=sr_a, target_sr=target_sr)
if sr_b != target_sr: b = librosa.resample(b, orig_sr=sr_b, target_sr=target_sr)

# pad to same length
max_len = max(len(a), len(b))
a = np.pad(a, (0, max_len - len(a)))
b = np.pad(b, (0, max_len - len(b)))

# chunk & compare
chunk_samples = int(target_sr * chunk_ms / 1000)
n_chunks = max_len // chunk_samples
trim = n_chunks * chunk_samples
a = a[:trim].reshape(n_chunks, chunk_samples)
b = b[:trim].reshape(n_chunks, chunk_samples)

diff_rms = np.sqrt(np.mean((a - b) ** 2, axis=1))
max_rms = np.maximum(np.sqrt(np.mean(a**2, axis=1)), np.sqrt(np.mean(b**2, axis=1))) + 1e-10
diff_mask = (diff_rms > threshold) | (diff_rms / max_rms > 0.5)

# merge nearby diffs into regions
diff_idx = np.where(diff_mask)[0]
min_gap = int(min_gap_ms / chunk_ms)
regions = []
if len(diff_idx):
    start = end = diff_idx[0]
    for i in diff_idx[1:]:
        if i - end <= min_gap:
            end = i
        else:
            regions.append((start * chunk_ms / 1000, (end + 1) * chunk_ms / 1000))
            start = end = i
    regions.append((start * chunk_ms / 1000, (end + 1) * chunk_ms / 1000))

for s, e in regions:
    print(f"{s:.3f} - {e:.3f}  ({e-s:.3f}s)")