File size: 3,942 Bytes
48b78bf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | """Recompute per-clip CER and SECS for a folder of released clips.
Usage:
python score_released.py --wavs stress_set/audio --manifest stress_set/manifest.tsv \
--reference voices/reference_merged.wav --out stress_set/scores.csv [--no-cer] [--no-secs]
CER : hishab/titu_stt_bn_conformer_large (NeMo) transcripts vs the manifest text, both passed through the
same Stage-2 normalisation used at synthesis (bn_norm.for_cer: digits->words, honorific expansion,
unicode normalisation, punctuation/whitespace/zero-width removal). Character-level Levenshtein / len(ref).
SECS: ECAPA-TDNN (speechbrain/spkrec-ecapa-voxceleb) cosine similarity between the clip and the reference
voice, both resampled to 16 kHz.
manifest.tsv columns: id <TAB> text <TAB> group. Clips are <id>.wav.
Requires: bn_norm.py next to this script (shipped in the model repo, inference/bn_norm.py).
"""
import os, sys, glob, csv, argparse, numpy as np
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
def cer(r, h):
d = np.zeros((len(r) + 1, len(h) + 1), int)
for i in range(len(r) + 1): d[i][0] = i
for j in range(len(h) + 1): d[0][j] = j
for i in range(1, len(r) + 1):
for j in range(1, len(h) + 1): d[i][j] = min(d[i-1][j] + 1, d[i][j-1] + 1, d[i-1][j-1] + (r[i-1] != h[j-1]))
return d[len(r)][len(h)] / max(1, len(r))
def main():
ap = argparse.ArgumentParser(); ap.add_argument("--wavs", required=True); ap.add_argument("--manifest", required=True)
ap.add_argument("--reference", default=None); ap.add_argument("--out", required=True)
ap.add_argument("--no-cer", action="store_true"); ap.add_argument("--no-secs", action="store_true"); a = ap.parse_args()
rows = {l.split("\t")[0]: l.rstrip("\n").split("\t") for l in open(a.manifest, encoding="utf-8") if "\t" in l}
files = sorted(glob.glob(os.path.join(a.wavs, "*.wav"))); ids = [os.path.basename(p)[:-4] for p in files]
hyps, cers, secs = [""] * len(files), [""] * len(files), [""] * len(files)
if not a.no_cer:
import torch, nemo.collections.asr as nemo_asr
from bn_norm import for_cer as norm
m = nemo_asr.models.ASRModel.from_pretrained("hishab/titu_stt_bn_conformer_large", map_location="cuda" if torch.cuda.is_available() else "cpu").eval()
hyps = [(o.text if hasattr(o, "text") else o) for o in m.transcribe(files, batch_size=16)]
cers = [f"{cer(norm(rows[i][1]), norm(h)):.4f}" for i, h in zip(ids, hyps)]
if not a.no_secs:
import torch, torchaudio
from speechbrain.inference.speaker import EncoderClassifier
dev = "cuda" if torch.cuda.is_available() else "cpu"
enc = EncoderClassifier.from_hparams(source="speechbrain/spkrec-ecapa-voxceleb", run_opts={"device": dev})
def emb(p):
w, sr = torchaudio.load(p); w = w.mean(0, keepdim=True)
if sr != 16000: w = torchaudio.functional.resample(w, sr, 16000)
with torch.no_grad(): return enc.encode_batch(w.to(dev)).squeeze().cpu().numpy()
ref = emb(a.reference)
secs = []
for p in files:
e = emb(p); secs.append(f"{float(e @ ref / (np.linalg.norm(e) * np.linalg.norm(ref))):.4f}")
with open(a.out, "w", encoding="utf-8", newline="") as f:
w = csv.writer(f); w.writerow(["id", "group", "text", "asr_hypothesis", "cer", "secs"])
for i, h, c, s in zip(ids, hyps, cers, secs): w.writerow([i, rows[i][2], rows[i][1], h, c, s])
import collections
g = collections.defaultdict(lambda: [[], []])
for i, c, s in zip(ids, cers, secs):
if c: g[rows[i][2]][0].append(float(c))
if s: g[rows[i][2]][1].append(float(s))
for k in sorted(g):
c, s = g[k]
print(f"{k:28s} n={max(len(c), len(s)):3d}" + (f" CER mean={100*np.mean(c):5.2f}% median={100*np.median(c):5.2f}%" if c else "") + (f" SECS={np.mean(s):.3f}" if s else ""))
if __name__ == "__main__": main()
|