Datasets:
File size: 3,896 Bytes
434a66e | 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 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 | #!/usr/bin/env python3
"""Score STT transcripts of your TTS outputs against TR-TTS-Bench (stdlib only; no GPU needed).
python score_transcripts.py --segments data/public/segments.jsonl --hyps my_system_R.jsonl --track R
`--hyps` is JSON Lines with {"id": "trtb-00001", "hyp": "<STT transcript of your audio for that segment>"}.
Segments without a transcript count as failures (all deletions). Track R = your TTS was given `text`; track S = it
was given `text_spoken` (scored strictly against the canonical reading). Prints headline WER/CER/SER, entity
accuracy and per-group WER; `--out` writes per-segment scores.
"""
from __future__ import annotations
import argparse
import json
import sys
from collections import defaultdict
from pathlib import Path
HERE = Path(__file__).resolve().parent
for p in (HERE, HERE.parent): # works from the repo (scripts/) and from the HF snapshot (code/)
if (p / "trttsbench").is_dir():
sys.path.insert(0, str(p))
from trttsbench.normalize import normalize_hyp # noqa: E402
from trttsbench.score import score_segment # noqa: E402
from trttsbench.textnorm import parse_lattice # noqa: E402
def main() -> int:
ap = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
ap.add_argument("--segments", required=True)
ap.add_argument("--hyps", required=True)
ap.add_argument("--track", choices=["R", "S"], default="R")
ap.add_argument("--out", help="write per-segment scores (JSON Lines)")
a = ap.parse_args()
rows = [json.loads(x) for x in Path(a.segments).read_text(encoding="utf-8").splitlines() if x.strip()]
hyps = {}
for x in Path(a.hyps).read_text(encoding="utf-8").splitlines():
if x.strip():
d = json.loads(x)
hyps[d["id"]] = d.get("hyp", "")
strict = a.track == "S"
tot = defaultdict(lambda: defaultdict(float))
per = []
for r in rows:
failed = r["id"] not in hyps
s = score_segment(parse_lattice(r["ref_lattice"]), normalize_hyp(hyps.get(r["id"], "")), failed=failed,
strict=strict)
s.update(id=r["id"], group=r["group"], headline=r["headline"], long_form=r.get("long_form", False))
per.append(s)
scopes = ["headline" if r["headline"] else ("long_form" if r.get("long_form") else "diagnostic")]
if r["headline"]:
scopes.append(f"group:{r['group']} {r['group_name']}")
for sc in scopes:
t = tot[sc]
t["n"] += 1
t["err"] += s["errors"]
t["ref"] += s["ref_words"]
t["cerr"] += s["char_errors"]
t["cref"] += s["ref_chars"]
t["sent"] += bool(s["sentence_error"])
t["ent"] += s["n_entities"]
t["entok"] += s["n_entities_correct"]
t["fail"] += failed
def line(name: str, t: dict) -> str:
wer = 100 * t["err"] / max(1, t["ref"])
cer = 100 * t["cerr"] / max(1, t["cref"])
ser = 100 * t["sent"] / max(1, t["n"])
ent = f"{100 * t['entok'] / t['ent']:.1f}" if t["ent"] else "–"
return f"| {name} | {int(t['n'])} | {wer:.2f} | {cer:.2f} | {ser:.1f} | {ent} | {int(t['fail'])} |"
print(f"Track {a.track} ({'strict canonical' if strict else 'lattice: min over accepted readings'}), "
f"{len(hyps)} transcripts for {len(rows)} segments\n")
print("| scope | n | WER % | CER % | SER % | entity acc % | missing |\n|---|---:|---:|---:|---:|---:|---:|")
for k in ["headline", "long_form", "diagnostic"] + sorted(k for k in tot if k.startswith("group:")):
if k in tot:
print(line(k, tot[k]))
if a.out:
with open(a.out, "w", encoding="utf-8") as f:
for s in per:
f.write(json.dumps(s, ensure_ascii=False) + "\n")
return 0
if __name__ == "__main__":
raise SystemExit(main())
|