#!/usr/bin/env python3 """ ECAPA-TDNN 声纹搜索 — 多线程版本。 模型已缓存在 ~/.cache/huggingface，不再重复下载。 """ import os, sys, shutil, json, time, logging from multiprocessing import Pool, cpu_count import numpy as np import torch import torchaudio SEARCH_DIR = r"D:\Project\Code\Uni\Cyrene-Voice-Model\data\cleaned" OUT_DIR = r"D:\Project\Code\Uni\Cyrene-Voice-Model\data\cyrene_ecapa" WORKERS = max(1, cpu_count() - 1) os.environ.setdefault("HF_ENDPOINT", "https://hf-mirror.com") CONFIRMED = [ "External_del_3.5_chapter_2/External_del_3.5_chapter_2_0358_02ff22a9.wav", "External_del_3.5_chapter_2/External_del_3.5_chapter_2_0369_0315d90a.wav", "External_del_3.5_chapter_2/External_del_3.5_chapter_2_0370_0317b5bf.wav", "External_del_3.5_chapter_2/External_del_3.5_chapter_2_0391_033cbeea.wav", "External_del_3.5_chapter_2/External_del_3.5_chapter_2_0392_033d9e2e.wav", "External_del_3.5_chapter_2/External_del_3.5_chapter_2_0398_0346fa30.wav", "External_del_3.5_chapter_2/External_del_3.5_chapter_2_0417_0367cbc5.wav", "External_del_3.5_chapter_2/External_del_3.5_chapter_2_0330_02cb864d.wav", "External_del_3.5_chapter_2/External_del_3.5_chapter_2_0357_02fd4ab1.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0106_009dfe25.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0107_009fba53.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0112_00a89eb2.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0124_00c08c68.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0128_00c6ba19.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0365_026cf0de.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0371_0278f79f.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0374_02803ab5.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0409_02bcd547.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0410_02bec9f4.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0437_02e57ff0.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0440_02e7d98f.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0467_031daa9e.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0102_0095ba46.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0109_00a458cc.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0126_00c44990.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0130_00c95298.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0359_02670f55.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0364_026b8023.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0366_026f3922.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0369_0274d62a.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0372_027b67ef.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0392_029e92a6.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0413_02c1a418.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0415_02c466f1.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0417_02c8ffc0.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0430_02dbad24.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0431_02dda165.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0434_02e216cf.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0443_02eb0bc4.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0455_03048430.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0462_0316bcd9.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0463_03181c56.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0464_03195142.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0468_031e8bd6.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0469_03208c8a.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0484_03339e50.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0488_033a7664.wav", "External_del_3.4_chapter_0/External_del_3.4_chapter_0_0489_033b5c44.wav", ] def load_audio(path): wav, sr = torchaudio.load(path) if sr != 16000: wav = torchaudio.functional.resample(wav, sr, 16000) if wav.shape[0] > 1: wav = wav.mean(dim=0, keepdim=True) return wav # [1, samples] def get_ref_embeddings(): """主进程: 加载模型, 提取参考嵌入""" from speechbrain.inference.speaker import EncoderClassifier classifier = EncoderClassifier.from_hparams( source="speechbrain/spkrec-ecapa-voxceleb", run_opts={"device": "cpu"}, ) refs = [] for fname in CONFIRMED: path = os.path.join(SEARCH_DIR, fname) wav = load_audio(path) wav = wav.squeeze(0).unsqueeze(0) # [1, time] if wav.shape[1] < 16000: wav = torch.nn.functional.pad(wav, (0, 16000 - wav.shape[1])) with torch.no_grad(): emb = classifier.encode_batch(wav).squeeze() refs.append(emb.numpy()) return np.mean(refs, axis=0).astype(np.float32) def scan_chunk(args): """Worker: 加载模型, 扫描一批文件, 返回 [(sim, path), ...]""" paths, template_arr = args from speechbrain.inference.speaker import EncoderClassifier classifier = EncoderClassifier.from_hparams( source="speechbrain/spkrec-ecapa-voxceleb", run_opts={"device": "cpu"}, ) template = torch.from_numpy(template_arr) results = [] for path in paths: try: wav = load_audio(path) wav = wav.squeeze(0).unsqueeze(0) if wav.shape[1] < 8000: continue with torch.no_grad(): emb = classifier.encode_batch(wav).squeeze() sim = torch.nn.functional.cosine_similarity(emb, template, dim=0).item() results.append((sim, path)) except: pass return results def main(): os.makedirs(OUT_DIR, exist_ok=True) LOG_FILE = os.path.join(OUT_DIR, "search.log") logging.basicConfig( level=logging.INFO, format="%(asctime)s %(message)s", datefmt="%H:%M:%S", handlers=[logging.FileHandler(LOG_FILE, encoding='utf-8'), logging.StreamHandler(sys.stdout)], ) log = logging.getLogger("ecapa") log.info("ECAPA-TDNN Search (multiprocess)") log.info(f" refs: {len(CONFIRMED)} | workers: {WORKERS}") # Step 1: 参考嵌入 (主进程) log.info("Extracting reference embeddings...") t0 = time.time() template = get_ref_embeddings() log.info(f" template: dim={len(template)}, pitch proxy={template[0]:.4f} ({time.time()-t0:.0f}s)") # Step 2: 收集文件 all_wavs = [] for root, dirs, files in os.walk(SEARCH_DIR): for f in files: if f.endswith(".wav"): all_wavs.append(os.path.join(root, f)) log.info(f" files: {len(all_wavs):,}") # Step 3: 分块, 多线程扫描 chunk_size = max(50, len(all_wavs) // (WORKERS * 4)) chunks = [all_wavs[i:i+chunk_size] for i in range(0, len(all_wavs), chunk_size)] chunk_args = [(chunk, template) for chunk in chunks] log.info(f" chunks: {len(chunks)} x ~{chunk_size} | starting pool...") t1 = time.time() results = [] done = 0 pool = Pool(WORKERS) for chunk_results in pool.imap_unordered(scan_chunk, chunk_args): results.extend(chunk_results) done += chunk_size elapsed = time.time() - t1 rate = min(done, len(all_wavs)) / elapsed if elapsed else 0 eta = (len(all_wavs) - min(done, len(all_wavs))) / rate if rate else 0 pct = min(done, len(all_wavs)) * 100 / len(all_wavs) print(f" [{pct:5.1f}%] {min(done, len(all_wavs)):,}/{len(all_wavs):,} | " f"{rate:.0f} f/s | ETA {eta:.0f}s | {len(results):,} ok") pool.close() pool.join() print() log.info(f" scanned in {time.time()-t1:.0f}s | {len(results):,} results") # Step 4: 排序 + 输出 results.sort(key=lambda x: x[0], reverse=True) log.info(f"\n{'='*55}") log.info("Top 50 Candidates") log.info("=" * 55) for rank, (sim, path) in enumerate(results[:50], 1): d = os.path.basename(os.path.dirname(path)) f = os.path.basename(path) log.info(f" {rank:2d}. [{sim:.4f}] {d}/{f}") log.info(f"\nSource distribution (sim > 0.65):") srcs = {} for sim, path in results: if sim > 0.65: d = os.path.basename(os.path.dirname(path)) srcs[d] = srcs.get(d, 0) + 1 for d in sorted(srcs): log.info(f" {d}: {srcs[d]}") # Export tiers = [("tier1_075", 0.75), ("tier2_070", 0.70), ("tier3_065", 0.65), ("tier4_060", 0.60)] for tier_name, thresh in tiers: tier_dir = os.path.join(OUT_DIR, tier_name) os.makedirs(tier_dir, exist_ok=True) n = 0 for sim, path in results: if sim >= thresh: dst = os.path.join(tier_dir, os.path.basename(path)) if os.path.exists(path) and not os.path.exists(dst): shutil.copy2(path, dst) n += 1 log.info(f" {tier_name}: {n} files") rp = os.path.join(OUT_DIR, "results.json") with open(rp, 'w') as f: json.dump([(float(s), p) for s, p in results], f, ensure_ascii=False) log.info(f"\n DONE ({time.time()-t0:.0f}s) | {rp}") if __name__ == "__main__": main()