| """ |
| ViuAI_TTS_Dataset Fix Script — Kaggle par chalao (local kuch nahi). |
| |
| Problem: HF repo me audio/ me orphan WAVs hain jinka metadata.csv me |
| koi transcript/row nahi (purane collector bug se). Training unko ignore |
| karta hai, repo ganda rehta hai. |
| |
| Usage (Kaggle single cell): |
| !pip install -q datasets huggingface_hub soundfile |
| !wget -q https://huggingface.co/datasets/ViuAI/ViuAI_TTS_Dataset/resolve/main/kaggle_fix_dataset.py |
| !python kaggle_fix_dataset.py --mode all |
| |
| Ya step-by-step: --mode audit | fix | push |
| |
| Token: Kaggle Secrets me HF_TOKEN (Write access) add karo, ya --token do. |
| """ |
|
|
| import os |
| import sys |
| import csv |
| import glob |
| import shutil |
| import argparse |
| from collections import Counter |
|
|
| if hasattr(sys.stdout, "reconfigure"): |
| try: |
| sys.stdout.reconfigure(encoding="utf-8") |
| except Exception: |
| pass |
|
|
|
|
| def get_token(args_token=None): |
| token = args_token or os.environ.get("HF_TOKEN") |
| if not token: |
| try: |
| from kaggle_secrets import UserSecretsClient |
| token = UserSecretsClient().get_secret("HF_TOKEN") |
| except Exception: |
| pass |
| if not token: |
| from huggingface_hub import get_token as hf_get_token |
| token = hf_get_token() |
| if not token: |
| raise SystemExit("[!] HF token nahi mila. Kaggle Secrets me HF_TOKEN add karo ya --token do.") |
| return token |
|
|
|
|
| def download_repo(repo, workdir, token): |
| from huggingface_hub import snapshot_download, login |
| login(token=token) |
| snapshot_download(repo_id=repo, repo_type="dataset", local_dir=workdir) |
| print(f"[+] Downloaded {repo} -> {workdir}") |
|
|
|
|
| def load_state(workdir): |
| wavs = {os.path.basename(p) for p in glob.glob(os.path.join(workdir, "audio", "*.wav"))} |
| rows = [] |
| with open(os.path.join(workdir, "metadata.csv"), encoding="utf-8") as f: |
| for row in csv.reader(f, delimiter="|"): |
| if len(row) >= 4 and row[0] != "audio_file": |
| rows.append(row) |
| return wavs, rows |
|
|
|
|
| def do_audit(workdir): |
| wavs, rows = load_state(workdir) |
| meta_files = {os.path.basename(r[0]) for r in rows} |
| orphans = sorted(wavs - meta_files) |
| missing = sorted(meta_files - wavs) |
| print("=" * 60) |
| print(f"Disk WAVs : {len(wavs)}") |
| print(f"Metadata rows : {len(rows)}") |
| print(f"Orphan WAVs : {len(orphans)} (transcript nahi, training me bekaar)") |
| print(f"Missing files : {len(missing)} (row hai, audio nahi)") |
| print(f"By language : {dict(Counter(r[3].strip() for r in rows))}") |
| print(f"Speakers : {sorted({r[2].strip() for r in rows})}") |
| print("Sample orphans :", orphans[:5]) |
| print("Sample missing :", missing[:5]) |
| print("=" * 60) |
|
|
|
|
| def do_fix(workdir): |
| wavs, rows = load_state(workdir) |
| meta_files = {os.path.basename(r[0]) for r in rows} |
| orphans = sorted(wavs - meta_files) |
| qdir = "./quarantine" |
| os.makedirs(qdir, exist_ok=True) |
| for fn in orphans: |
| shutil.move(os.path.join(workdir, "audio", fn), os.path.join(qdir, fn)) |
| disk_now = {os.path.basename(p) for p in glob.glob(os.path.join(workdir, "audio", "*.wav"))} |
| final = [r for r in rows if os.path.basename(r[0]) in disk_now] |
| with open(os.path.join(workdir, "metadata.csv"), "w", encoding="utf-8", newline="") as f: |
| w = csv.writer(f, delimiter="|") |
| w.writerow(["audio_file", "transcript", "speaker", "language"]) |
| w.writerows(final) |
| print(f"[+] Quarantined orphans : {len(orphans)} (./quarantine, session ke sath discard)") |
| print(f"[+] Dropped stale rows : {len(rows) - len(final)}") |
| print(f"[+] FINAL clips : {len(final)} (WAVs == rows, 1:1)") |
| print(f" By language: {dict(Counter(r[3].strip() for r in final))}") |
|
|
|
|
| def do_push(repo, workdir, token): |
| from huggingface_hub import HfApi |
| api = HfApi(token=token) |
| remote_files = api.list_repo_files(repo_id=repo, repo_type="dataset") |
| remote_wavs = [f for f in remote_files if f.startswith("audio/") and f.endswith(".wav")] |
| _, rows = load_state(workdir) |
| keep = {r[0].strip() for r in rows} |
| to_delete = [f for f in remote_wavs if f not in keep] |
| print(f"Remote WAVs: {len(remote_wavs)}, keep: {len(keep)}, delete: {len(to_delete)}") |
| if not rows: |
| raise SystemExit("[!] metadata.csv khaali hai — push roka (pehle --mode fix chalao).") |
| for i in range(0, len(to_delete), 200): |
| batch = to_delete[i:i + 200] |
| api.delete_files(repo_id=repo, repo_type="dataset", delete_patterns=batch, |
| commit_message=f"Remove {len(batch)} orphan WAVs (no transcript)") |
| print(f" deleted batch {i // 200 + 1}: {len(batch)} files") |
| api.upload_file(path_or_fileobj=os.path.join(workdir, "metadata.csv"), |
| path_in_repo="metadata.csv", repo_id=repo, repo_type="dataset", |
| commit_message=f"Clean metadata: {len(keep)} clips, WAVs == rows 1:1") |
| print(f"[DONE] Verify: https://huggingface.co/datasets/{repo}") |
|
|
|
|
| def fresh_download(repo, workdir, token): |
| """Stale-session guard: purana workdir mitao, HF se taaza snapshot lao.""" |
| if os.path.exists(workdir): |
| import shutil |
| shutil.rmtree(workdir) |
| print(f"[*] Stale local copy hatayi: {workdir}") |
| download_repo(repo, workdir, token) |
|
|
|
|
| def do_all(repo, workdir, token): |
| """Single-cell full run: download -> audit -> fix -> verify 1:1 -> push.""" |
| fresh_download(repo, workdir, token) |
| print("\n### STEP 1: AUDIT (before)\n") |
| do_audit(workdir) |
| print("\n### STEP 2: FIX\n") |
| do_fix(workdir) |
| |
| wavs, rows = load_state(workdir) |
| disk = {os.path.basename(p) for p in glob.glob(os.path.join(workdir, "audio", "*.wav"))} |
| meta = {os.path.basename(r[0]) for r in rows} |
| print("\n### STEP 3: VERIFY GATE\n") |
| print(f"Disk WAVs: {len(disk)}, Metadata rows: {len(rows)}") |
| if disk != meta: |
| raise SystemExit(f"[!] 1:1 match nahi hua (diff {len(disk ^ meta)}). Push ROKA. Audit dekho.") |
| print("[OK] 1:1 match — push kar rahe hain...\n") |
| print("### STEP 4: PUSH\n") |
| do_push(repo, workdir, token) |
|
|
|
|
| def qc_file(path): |
| """ |
| Real-human audio QC. Returns (ok: bool, reason: str). |
| Rejects: unreadable, <1s, >20s, silent (RMS<0.01), clipped (>5% samples), |
| wrong channels handled (mixed to mono at train time). |
| NOTE: sources (IndicTTS studio + LibriSpeech audiobook) are all real human |
| recordings — koi TTS/robot data upstream me nahi hai. Yeh QC garbage |
| (silence/corrupt/clipped) hatata hai, robot-tone detect karna iska kaam nahi. |
| """ |
| import numpy as np |
| try: |
| import soundfile as sf |
| data, sr = sf.read(path, dtype="float32", always_2d=True) |
| except Exception as e: |
| return False, f"unreadable ({e})" |
| dur = len(data) / float(sr) |
| if dur < 1.0: |
| return False, f"too_short ({dur:.2f}s)" |
| if dur > 20.0: |
| return False, f"too_long ({dur:.2f}s)" |
| mono = data.mean(axis=1) |
| rms = float((mono ** 2).mean() ** 0.5) |
| if rms < 0.01: |
| return False, f"silent (rms={rms:.4f})" |
| clip_frac = float((abs(mono) > 0.99).mean()) |
| if clip_frac > 0.05: |
| return False, f"clipped ({clip_frac * 100:.1f}%)" |
| return True, "ok" |
|
|
|
|
| def do_clean(repo, workdir, token): |
| """Full garbage-removal pipeline: download -> orphan fix -> audio QC -> verify -> push.""" |
| fresh_download(repo, workdir, token) |
| print("\n### STEP 1: ORPHAN FIX\n") |
| do_fix(workdir) |
| print("\n### STEP 2: AUDIO QC (har clip sunke nahi, naap ke — 5k files, kuch min)\n") |
| _, rows = load_state(workdir) |
| qdir = "./quarantine" |
| os.makedirs(qdir, exist_ok=True) |
| q_reasons = Counter() |
| kept = [] |
| for i, r in enumerate(rows): |
| rel = r[0].strip() |
| p = os.path.join(workdir, rel) |
| if not os.path.exists(p): |
| q_reasons["missing_file"] += 1 |
| continue |
| ok, reason = qc_file(p) |
| if ok: |
| kept.append(r) |
| else: |
| q_reasons[reason.split("(")[0].strip()] += 1 |
| shutil.move(p, os.path.join(qdir, os.path.basename(p))) |
| if (i + 1) % 1000 == 0: |
| print(f" checked {i + 1}/{len(rows)}...") |
| print(f"[+] QC pass: {len(kept)}/{len(rows)} | quarantined: {dict(q_reasons)}") |
| with open(os.path.join(workdir, "metadata.csv"), "w", encoding="utf-8", newline="") as f: |
| w = csv.writer(f, delimiter="|") |
| w.writerow(["audio_file", "transcript", "speaker", "language"]) |
| w.writerows(kept) |
| print("\n### STEP 3: VERIFY GATE\n") |
| disk = {os.path.basename(p) for p in glob.glob(os.path.join(workdir, "audio", "*.wav"))} |
| meta = {os.path.basename(r[0]) for r in kept} |
| print(f"Disk WAVs: {len(disk)}, Metadata rows: {len(kept)}") |
| if disk != meta: |
| raise SystemExit("[!] 1:1 match nahi hua. Push ROKA.") |
| print("[OK] 1:1 match — push kar rahe hain...\n") |
| print("### STEP 4: PUSH\n") |
| do_push(repo, workdir, token) |
|
|
|
|
| if __name__ == "__main__": |
| p = argparse.ArgumentParser(description="ViuAI_TTS_Dataset fix (Kaggle)") |
| p.add_argument("--mode", choices=["audit", "fix", "push", "all", "clean"], default="audit", |
| help="all = orphan fix+push | clean = orphan fix + audio QC + push (single cell)") |
| p.add_argument("--repo", default="ViuAI/ViuAI_TTS_Dataset") |
| p.add_argument("--workdir", default="./hf_dataset") |
| p.add_argument("--token", default=None) |
| a = p.parse_args() |
|
|
| tok = get_token(a.token) |
| if a.mode == "all": |
| do_all(a.repo, a.workdir, tok) |
| raise SystemExit(0) |
| if a.mode == "clean": |
| do_clean(a.repo, a.workdir, tok) |
| raise SystemExit(0) |
| if a.mode in ("audit", "fix") and not os.path.exists(os.path.join(a.workdir, "metadata.csv")): |
| download_repo(a.repo, a.workdir, tok) |
| if a.mode == "audit": |
| do_audit(a.workdir) |
| elif a.mode == "fix": |
| do_fix(a.workdir) |
| elif a.mode == "push": |
| do_push(a.repo, a.workdir, tok) |
|
|
