feat: overlap-window speechbrain enhance, hop 200ms / window 1s

scripts/bench_enhance.py

@@ -0,0 +1,23 @@
+#!/usr/bin/env python3
+"""RTF bench for hop/window speechbrain enhance."""
+import time
+
+import numpy as np
+
+from audio_cleanup import AudioCleaner
+
+c = AudioCleaner(mode="force", chunk_s=1.0, hop_s=0.2)
+hop = 3200
+pcm = (np.random.randn(hop) * 800).astype(np.int16)
+for _ in range(8):
+    c.process(pcm)
+n = 50
+t0 = time.perf_counter()
+for _ in range(n):
+    c.process(pcm)
+dt = time.perf_counter() - t0
+audio_s = n * 0.2
+rtf = dt / audio_s
+print(f"hops={n} wall={dt:.3f}s audio={audio_s:.1f}s rtf={rtf:.3f}")
+if rtf > 0.8:
+    raise SystemExit("RTF too high for 20-way realtime")