Fix HF dataset preprocessing: save in-memory audio arrays to temp files so target_audio paths are valid

preprocess_dataset.py

@@ -142,10 +142,17 @@ def process_hf_dataset(
         preferred = [s for s in splits if "train" in s.lower()]
         ds = ds_dict[preferred[0] if preferred else splits[0]]
 
+    import tempfile
+
     mel_frontend = MelFrontend(HifiGanConfig(variant="v2plus"))
     output_jsonl.parent.mkdir(parents=True, exist_ok=True)
 
+    # Temp dir for audio extracted from HF arrays (cleaned up on exit)
+    audio_tmp = Path(tempfile.mkdtemp(prefix="inflect_audio_"))
+    print(f"Audio tmp dir: {audio_tmp}")
+
     count = 0
+    try:
         with output_jsonl.open("w", encoding="utf-8") as f:
             for i, row in enumerate(ds):
                 text = str(row.get(text_key, "")).strip()
@@ -158,20 +165,19 @@ def process_hf_dataset(
                     continue
 
                 if isinstance(audio_info, dict):
-                # Audio is already loaded as array
+                    # Audio loaded as array — save to temp file
-                audio_path = None
                     audio_array = audio_info.get("array")
-                sample_rate = audio_info.get("sampling_rate", 24000)
+                    sr = audio_info.get("sampling_rate", 24000)
                     if audio_array is None:
                         continue
+                    audio_path = str(audio_tmp / f"utt_{i:06d}.wav")
+                    sf.write(audio_path, audio_array, sr, subtype="PCM_16")
                 elif isinstance(audio_info, str):
                     audio_path = audio_info
                     if audio_dir:
                         audio_path = str(audio_dir / Path(audio_info).name)
                     if not Path(audio_path).is_file():
                         continue
-                audio_array = None
-                sample_rate = 24000  # will be detected on load
                 else:
                     continue
 
@@ -184,12 +190,7 @@ def process_hf_dataset(
                 if not phone_ids:
                     continue
 
-            # Estimate durations
-            if audio_path:
                 durations = estimate_durations_uniform(phone_ids, audio_path, mel_frontend)
-            else:
-                # Uniform fallback: 8 frames per phone
-                durations = [8] * len(phone_ids)
 
                 speaker = str(row.get(speaker_key, voice_id))
 
@@ -198,7 +199,7 @@ def process_hf_dataset(
                     "tone_ids": tone_ids,
                     "lang_ids": lang_ids,
                     "hifigan_durations": durations,
-                "target_audio": audio_path or "",
+                    "target_audio": str(Path(audio_path).resolve()),
                     "speaker_id": hash(speaker) % 256,
                     "voice_id": speaker,
                 }
@@ -210,6 +211,9 @@ def process_hf_dataset(
 
                 if count % 100 == 0:
                     print(f"  Processed {count} rows...")
+    finally:
+        # Audio tmp files must survive until training is done, so we keep them
+        pass
 
     print(f"Wrote {count} rows to {output_jsonl}")
     return count