| """ |
| Curry speaker verification: pair scoring and VoxCeleb1-O evaluation. |
| |
| python inference.py a.wav b.wav # cosine score for one pair |
| python inference.py --trials veri_test.txt --root /path/to/VoxCeleb1/wav/ # EER / minDCF |
| python inference.py --repo /local/dir ... # use a local copy instead of the Hub |
| """ |
| import argparse |
| import numpy as np |
| import torch |
| from tqdm import tqdm |
| from scipy.interpolate import interp1d |
| from scipy.optimize import brentq |
| from sklearn.metrics import roc_curve |
| from transformers import AutoModel |
|
|
| REPO = "cmu-mlsp/Curry" |
|
|
|
|
| def compute_eer(labels, scores): |
| fpr, tpr, thr = roc_curve(labels, scores, pos_label=1) |
| eer = brentq(lambda x: 1.0 - x - interp1d(fpr, tpr)(x), 0.0, 1.0) |
| return eer, float(interp1d(fpr, thr)(eer)) |
|
|
|
|
| def compute_min_dcf(labels, scores, p_target=0.01, c_miss=1, c_fa=1): |
| fpr, tpr, thr = roc_curve(labels, scores, pos_label=1) |
| fnr = 1.0 - tpr |
| c_det = c_miss * fnr * p_target + c_fa * fpr * (1 - p_target) |
| i = int(np.argmin(c_det)) |
| return c_det[i] / min(c_miss * p_target, c_fa * (1 - p_target)), thr[i] |
|
|
|
|
| def main(): |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--repo", default=REPO, help="Hub repo id or local directory") |
| ap.add_argument("--trials", help="trial file: label enroll_path test_path") |
| ap.add_argument("--root", default="", help="prefix prepended to trial paths") |
| ap.add_argument("--no-mean-norm", action="store_true", help="disable eval-set embedding mean subtraction") |
| ap.add_argument("wavs", nargs="*") |
| args = ap.parse_args() |
|
|
| device = "cuda" if torch.cuda.is_available() else "cpu" |
| model = AutoModel.from_pretrained(args.repo, trust_remote_code=True).eval().to(device) |
|
|
| if args.trials: |
| trials = np.loadtxt(args.trials, str) |
| paths = sorted(set(trials[:, 1]) | set(trials[:, 2])) |
| embs = {p: model.embed_file(args.root + p).cpu().numpy() for p in tqdm(paths, desc="embedding")} |
| if not args.no_mean_norm: |
| mu = np.mean(list(embs.values()), axis=0) |
| embs = {k: v - mu for k, v in embs.items()} |
| labels, scores = [], [] |
| for lab, a, b in trials: |
| ea, eb = embs[a], embs[b] |
| scores.append(float(ea @ eb / (np.linalg.norm(ea) * np.linalg.norm(eb)))) |
| labels.append(int(lab)) |
| eer, thr = compute_eer(labels, scores) |
| print(f"cosine EER: {eer*100:.2f}% (threshold {thr:.3f})") |
| for p in (0.01, 0.001): |
| dcf, _ = compute_min_dcf(labels, scores, p_target=p) |
| print(f"minDCF(p={p}): {dcf:.4f}") |
| elif len(args.wavs) == 2: |
| print(f"cosine similarity: {model.verify(*args.wavs):.4f}") |
| else: |
| for w in args.wavs: |
| print(w, tuple(model.embed_file(w).shape)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
