Fix: shard audio into speaker subdirectories to bypass Git 10k files limit

colab_download_and_push.py

@@ -143,13 +143,15 @@ def collect_fast(
                 if wav_bytes is None:
                     skipped += 1
                     continue
+                spk_dir = os.path.join(audio_dir, speaker)
+                os.makedirs(spk_dir, exist_ok=True)
                 filename = f"{speaker}_{count+1:05d}.wav"
-                file_path = os.path.join(audio_dir, filename)
+                file_path = os.path.join(spk_dir, filename)
                 if not os.path.exists(file_path):
                     with open(file_path, "wb") as f:
                         f.write(wav_bytes)
                 new_records.append({
-                    "audio_file": f"audio/{filename}",
+                    "audio_file": f"audio/{speaker}/{filename}",
                     "transcript": text.strip(),
                     "speaker": speaker,
                     "language": lang,
@@ -243,13 +245,15 @@ def collect_english(audio_dir, new_records, english_samples, en_male_per_speaker
                 if is_male:
                     male_counts[spk] = male_counts.get(spk, 0) + 1
                 en_count += 1
+                spk_dir = os.path.join(audio_dir, speaker_tag)
+                os.makedirs(spk_dir, exist_ok=True)
                 filename = f"english_{spk}_{en_count:06d}.wav"
-                file_path = os.path.join(audio_dir, filename)
+                file_path = os.path.join(spk_dir, filename)
                 if not os.path.exists(file_path):
                     with open(file_path, "wb") as f:
                         f.write(wav_bytes)
                 new_records.append({
-                    "audio_file": f"audio/{filename}",
+                    "audio_file": f"audio/{speaker_tag}/{filename}",
                     "transcript": text.strip(),
                     "speaker": speaker_tag,
                     "language": "en",
@@ -303,15 +307,28 @@ def merge_and_upload(output_dir, metadata_path, new_records, token, full_target_
     print(f"  [*] By language: {dict(by_lang)}")
     print(f"  [*] Speakers ({len(by_spk)}): {sorted(by_spk)}")
 
-    # 4. Upload directly to Hugging Face
-    print(f"\n[+] Uploading combined dataset ({len(final_records)} clips) to Hugging Face: {full_target_repo}...")
+    # 4. Clean up legacy flat root audio files on Hugging Face (count('/') == 1) to prevent 10k directory limit
     api = HfApi(token=token)
+    try:
+        remote_files = api.list_repo_files(repo_id=full_target_repo, repo_type="dataset")
+        old_root_wavs = [f for f in remote_files if f.startswith("audio/") and f.count("/") == 1 and f.endswith(".wav")]
+        if old_root_wavs:
+            print(f"\n[+] Cleaning up {len(old_root_wavs)} legacy root audio files to prevent HF 10k directory limit...")
+            for i in range(0, len(old_root_wavs), 500):
+                batch = old_root_wavs[i:i + 500]
+                api.delete_files(repo_id=full_target_repo, repo_type="dataset", delete_patterns=batch,
+                                 commit_message=f"Remove legacy flat audio batch {i//500 + 1}")
+    except Exception as e:
+        print(f"  [!] Note on remote cleanup: {e}")
+
+    # 5. Upload directly to Hugging Face
+    print(f"\n[+] Uploading combined dataset ({len(final_records)} clips) to Hugging Face: {full_target_repo}...")
     api.upload_folder(
         folder_path=output_dir,
         repo_id=full_target_repo,
         repo_type="dataset",
         ignore_patterns=["*.tar.bz2", "LJSpeech*"],
-        commit_message=f"Dataset expansion: {len(final_records)} clips ({dict(by_lang)})",
+        commit_message=f"Dataset expansion (speaker-sharded): {len(final_records)} clips ({dict(by_lang)})",
     )
 
     print("\n" + "=" * 70)