whisper lines capped at 18 chars; retake short takes

kaggle/radio_demo.py

@@ -205,8 +205,12 @@ def main() -> None:
         for line in lines:
             v = voices[line["slot"]]
             wav_path = str(BUNDLE / v["wav"])
-            t0 = time.time()
-            try:
+            frames = (frames_by_index or {}).get(line["index"])
+            if use_student and not frames:
+                log(f"{tag} line {line['index']} skipped: no teacher frames")
+                continue
+
+            def gen(attempt: int) -> np.ndarray:
                 if use_student:
                     latent, spk = tts.prepare_prompt(wav_path, use_spk_emb=True)
                     latent = latent.to(device=device, dtype=dtype)
@@ -215,22 +219,33 @@ def main() -> None:
                     # teacher again to measure length gives a different take -
                     # 25 frames vs 16 on a whisper line - and a wrong budget
                     # makes the student drift into a loop or into silence.
-                    frames = (frames_by_index or {}).get(line["index"])
-                    if not frames:
-                        log(f"{tag} line {line['index']} skipped: no teacher frames")
-                        continue
-                    torch.manual_seed(SEED + line["index"])
-                    y = _ar(mm, sampler, line["text"], task_prompt, latent,
-                            v["prompt_text"], spk, frames)
-                else:
-                    y = tts.synthesize(task="TTS", text=line["text"], language_id="Chinese",
-                                       prompt_wav_path=wav_path, prompt_text=v["prompt_text"],
-                                       use_spk_emb=True, raw_text=True,
-                                       max_decode_steps=max(32, min(400, int(len(line["text"]) * 3.5))))
-                    y = y.cpu().numpy().reshape(-1)
+                    torch.manual_seed(SEED + 1000 * attempt + line["index"])
+                    return _ar(mm, sampler, line["text"], task_prompt, latent,
+                               v["prompt_text"], spk, frames)
+                out = tts.synthesize(task="TTS", text=line["text"], language_id="Chinese",
+                                     prompt_wav_path=wav_path, prompt_text=v["prompt_text"],
+                                     use_spk_emb=True, raw_text=True,
+                                     max_decode_steps=max(32, min(400, int(len(line["text"]) * 3.5))))
+                return out.cpu().numpy().reshape(-1)
+
+            t0 = time.time()
+            try:
+                y = gen(0)
             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")
+                try:
+                    y2 = gen(1)
+                    if y2.size > y.size:
+                        y = y2
+                except Exception:
+                    log(f"{tag} line {line['index']} retake failed (keeping the short take)")
             wall = time.time() - t0
             p = OUT / "lines" / f"{tag}_{line['index']:02d}.wav"
             sf.write(str(p), y, SR)