Spaces:
Paused
Paused
| #!/usr/bin/env python3 | |
| """ | |
| pick_voice_ref.py | |
| Pick the best reference clips for voice generation from the dataset made by prepare_dataset.py. | |
| Chatterbox copies the voice, tone and energy of ONE reference clip, using about the first | |
| 10 seconds of it. So the reference matters a lot: this script ranks the kept dataset clips | |
| by how strongly they match Brian's voice, how long they are, and how clean they are, then | |
| copies the top picks into a folder so you can audition them. | |
| Usage: | |
| python pick_voice_ref.py ..\\Brian-voice\\dataset_test --name brian | |
| python pick_voice_ref.py ..\\Lily-voice\\dataset --name lily | |
| """ | |
| import argparse | |
| import csv | |
| import re | |
| import shutil | |
| import sys | |
| from pathlib import Path | |
| import numpy as np | |
| import soundfile as sf | |
| def analyze(path): | |
| """Quick quality checks on a clip: loudness, silence, clipping.""" | |
| x, sr = sf.read(str(path), dtype="float32", always_2d=True) | |
| x = x.mean(axis=1) | |
| frame = int(sr * 0.05) | |
| n = len(x) // frame | |
| if n == 0: | |
| return None | |
| rms = np.sqrt(np.mean(x[: n * frame].reshape(n, frame) ** 2, axis=1) + 1e-12) | |
| db = 20 * np.log10(rms + 1e-12) | |
| speech = db > (db.max() - 35) | |
| return { | |
| "seconds": len(x) / sr, | |
| "speech_ratio": float(speech.mean()), # how much of the clip is talking | |
| "clipped": float(np.mean(np.abs(x) >= 0.999)), | |
| } | |
| def main(): | |
| p = argparse.ArgumentParser(description="Rank dataset clips as voice references.") | |
| p.add_argument("dataset", type=Path, help="Folder made by prepare_dataset.py (has report.csv)") | |
| p.add_argument("--name", default="brian", | |
| help="Voice name for the default output folder (e.g. lily -> lily_voice_refs)") | |
| p.add_argument("--out", type=Path, help="Output folder (default: <name>_voice_refs)") | |
| p.add_argument("--top", type=int, default=6, help="How many candidates to copy") | |
| p.add_argument("--min-sec", type=float, default=7.0, | |
| help="Shortest usable clip (Turbo needs more than 5 s)") | |
| p.add_argument("--max-sec", type=float, default=16.0) | |
| args = p.parse_args() | |
| voice_slug = re.sub(r"[^a-zA-Z0-9]+", "_", args.name.strip()).strip("_").lower() | |
| if not voice_slug: | |
| p.error("--name must contain at least one letter or number") | |
| out = args.out if args.out is not None else Path(f"{voice_slug}_voice_refs") | |
| report = args.dataset / "report.csv" | |
| wavs = args.dataset / "wavs" | |
| if not report.exists(): | |
| sys.exit(f"Can't find {report}. Point this at the folder prepare_dataset.py created.") | |
| rows = [] | |
| with open(report, newline="", encoding="utf-8") as f: | |
| for r in csv.DictReader(f): | |
| if r.get("status") != "kept": | |
| continue | |
| wav = wavs / f"{r['clip_id']}.wav" | |
| if not wav.exists(): | |
| continue | |
| q = analyze(wav) | |
| if q is None or not (args.min_sec <= q["seconds"] <= args.max_sec): | |
| continue | |
| if q["clipped"] > 0.0005: | |
| continue | |
| sim = float(r["similarity"]) | |
| # Strong voice match matters most; prefer clips that are mostly speech and ~10-14 s. | |
| length_bonus = 1.0 - min(abs(q["seconds"] - 12.0) / 6.0, 1.0) | |
| score = 3.0 * sim + 1.0 * q["speech_ratio"] + 0.5 * length_bonus | |
| rows.append({**r, **q, "wav": wav, "score": score}) | |
| if not rows: | |
| sys.exit(f"No kept clips between {args.min_sec:.0f} and {args.max_sec:.0f} seconds. " | |
| "Process more videos first, or lower --min-sec (not below 5.5).") | |
| rows.sort(key=lambda r: -r["score"]) | |
| out.mkdir(parents=True, exist_ok=True) | |
| print(f"{len(rows)} usable clips found. Top {min(args.top, len(rows))}:\n") | |
| with open(out / "refs.csv", "w", newline="", encoding="utf-8") as f: | |
| w = csv.writer(f) | |
| w.writerow(["file", "seconds", "similarity", "score", "source_clip", "text"]) | |
| for i, r in enumerate(rows[: args.top], 1): | |
| name = f"ref_{i:02d}.wav" | |
| shutil.copy2(r["wav"], out / name) | |
| w.writerow([name, f"{r['seconds']:.1f}", r["similarity"], f"{r['score']:.2f}", | |
| r["clip_id"], r["text"]]) | |
| print(f" {name} {r['seconds']:4.1f}s match {float(r['similarity']):.2f} " | |
| f"\"{r['text'][:70]}{'...' if len(r['text']) > 70 else ''}\"") | |
| print(f"\nCopied to {out.resolve()}") | |
| print("Listen to them and pick one whose energy fits the line you're generating:") | |
| print("calm explaining, excited, or talking to camera. That clip's tone carries over.") | |
| if __name__ == "__main__": | |
| main() | |