Datasets:
File size: 1,738 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 | """tr_norm v1 entry points (DESIGN.md §4).
Hypotheses (STT output) are verbalized with `trttsbench.verbalize` into a lattice of spoken alternatives and then
normalized with `textnorm.normalize_basic`; references are built from `text_spoken` + `alt_spans`. Both sides end
in the same token space, so `trttsbench.score.align(ref, hyp)` compares like with like.
"""
from __future__ import annotations
from .textnorm import TR_NORM_VERSION, Lattice, build_lattice, tokens
from .verbalize import verbalize_lattice
__all__ = ["TR_NORM_VERSION", "normalize_hyp", "normalize_hyp_str", "normalize_ref"]
def normalize_hyp(text: str) -> Lattice:
"""Lattice of normalized token sequences for a written (e.g. STT) string; first option = canonical."""
out: Lattice = []
for group in verbalize_lattice(text or ""):
opts: list[list[str]] = []
for o in group:
t = tokens(" ".join(o) if isinstance(o, (list, tuple)) else str(o))
if t not in opts:
opts.append(t)
if any(opts):
out.append([o for o in opts if o] if all(opts) else opts)
merged: Lattice = []
for g in out:
if len(g) == 1 and merged and len(merged[-1]) == 1:
merged[-1] = [merged[-1][0] + g[0]]
else:
merged.append(g)
return merged
def normalize_hyp_str(text: str) -> str:
"""Canonical normalized string of a hypothesis (first option of every group)."""
return " ".join(t for g in normalize_hyp(text) for t in g[0])
def normalize_ref(text_spoken: str, alt_spans: list[dict] | None = None) -> Lattice:
"""Reference lattice from the canonical spoken text and its span alternatives."""
return build_lattice(text_spoken, alt_spans)
|