lhotseWorkspace/cleanPoisonFiles.py

import os
import multiprocessing
from lhotse import CutSet
from tqdm.auto import tqdm

# --- EXACT PATHS ---
INPUT_MANIFEST = "/home/jonathan/lhotse_workspace/trimmed_utterance_cuts_16k_fixed.jsonl.gz"
OUTPUT_MANIFEST = "/home/jonathan/lhotse_workspace/bulletproof_cuts_for_shar.jsonl.gz"
POISON_REPORT = "/home/jonathan/lhotse_workspace/poison_files_report.txt"

def _isolated_audio_test(cut):
    """
    Runs in an isolated process. 
    If ffmpeg segfaults on a truncated file, it only kills this child process.
    """
    try:
        _ = cut.load_audio()
    except Exception:
        # Standard Python errors (e.g. file missing)
        os._exit(1)
    
    # Successfully read without dying
    os._exit(0)

def main():
    print(f"Loading master manifest: {INPUT_MANIFEST}")
    cuts = CutSet.from_file(INPUT_MANIFEST)
    
    # --- PHASE 1: Find the most dangerous cut for each physical file ---
    print("\n[Phase 1] Mapping physical files to find boundary cuts...")
    file_to_boundary_cut = {}
    
    for cut in tqdm(cuts, desc="Mapping"):
        if not cut.has_recording:
            continue
            
        source_path = cut.recording.sources[0].source
        
        # Test only the cut that reaches the FURTHEST into the audio file
        if source_path not in file_to_boundary_cut:
            file_to_boundary_cut[source_path] = cut
        else:
            if cut.end > file_to_boundary_cut[source_path].end:
                file_to_boundary_cut[source_path] = cut
                
    total_files = len(file_to_boundary_cut)
    print(f"Found {total_files} unique physical audio files.")
    
    # --- PHASE 2: Safely test the boundaries ---
    print("\n[Phase 2] Testing boundaries in isolated environments...")
    poison_files = set()
    
    for path, test_cut in tqdm(file_to_boundary_cut.items(), desc="Detonating Tests"):
        p = multiprocessing.Process(target=_isolated_audio_test, args=(test_cut,))
        p.start()
        p.join() # Wait for it to finish or crash
        
        # If exitcode != 0 (like -11 for Segfault), the file is poison.
        if p.exitcode != 0:
            poison_files.add(path)
            
    # --- PHASE 3: Amputation ---
    print(f"\n[Phase 3] Amputation. Found {len(poison_files)} poisoned files.")
    
    if len(poison_files) > 0:
        with open(POISON_REPORT, "w") as f:
            for pf in poison_files:
                f.write(f"{pf}\n")
        print(f"Saved corrupted file list to {POISON_REPORT}")
        
        # Filter out ANY cut that relies on a poisoned file
        print("Filtering corrupted files out of the manifest...")
        clean_cuts = cuts.filter(
            lambda c: c.has_recording and c.recording.sources[0].source not in poison_files
        )
    else:
        print("No poison files found! Your dataset is perfectly clean.")
        clean_cuts = cuts
        
    print(f"\nFinal clean manifest size: {len(clean_cuts)} cuts (Original: {len(cuts)})")
    print(f"Saving to {OUTPUT_MANIFEST}...")
    clean_cuts.to_file(OUTPUT_MANIFEST)
    print("\n✅ Done. You are clear to export.")

if __name__ == '__main__':
    # Required for safe multiprocessing in Python
    multiprocessing.set_start_method('spawn', force=True)
    main()