retake near-silent takes; whisper cap 15

kaggle/radio_demo.py

@@ -201,7 +201,7 @@ def main() -> None:
                     y = y + v * (1.0 / STEPS)
                 return y.detach()
 
-        pieces, rows = [], []
+        pieces, rows, peaks = [], [], []
         for line in lines:
             v = voices[line["slot"]]
             wav_path = str(BUNDLE / v["wav"])
@@ -234,18 +234,24 @@ def main() -> None:
             except Exception:
                 log(f"{tag} line {line['index']} failed\n{traceback.format_exc()}")
                 continue
-            # A collapsed take shows up as a fraction of the expected duration:
-            # one 24-character whisper line came out as 0.4 s of near-silence in
-            # both voices. One retake with a different seed is cheap and fixes
-            # most of them; keep whichever is longer.
-            if y.size / SR < 0.09 * len(line["text"]):
-                log(f"{tag} line {line['index']} short ({y.size/SR:.2f}s), retaking")
+            # A collapsed take shows up two ways: a fraction of the expected
+            # duration (a 24-character whisper line came out as 0.4 s), or the
+            # right duration filled with near-silence (peak 0.13 against 0.68
+            # for every other line). One retake with a different seed is cheap;
+            # keep whichever take is longer and louder.
+            ref_peak = float(np.median(peaks)) if len(peaks) >= 2 else 0.0
+            weak = float(np.abs(y).max()) < (0.45 * ref_peak if ref_peak else 0.20)
+            short = y.size / SR < 0.09 * len(line["text"])
+            if short or weak:
+                why = "short" if short else "near-silent"
+                log(f"{tag} line {line['index']} {why} ({y.size/SR:.2f}s peak {np.abs(y).max():.2f}), retaking")
                 try:
                     y2 = gen(1)
-                    if y2.size > y.size:
+                    if (y2.size, np.abs(y2).max()) > (y.size, np.abs(y).max()):
                         y = y2
                 except Exception:
-                    log(f"{tag} line {line['index']} retake failed (keeping the short take)")
+                    log(f"{tag} line {line['index']} retake failed (keeping the first take)")
+            peaks.append(float(np.abs(y).max()))
             wall = time.time() - t0
             p = OUT / "lines" / f"{tag}_{line['index']:02d}.wav"
             sf.write(str(p), y, SR)