File size: 1,316 Bytes
c202fc8 | 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 | """Text helpers shared by the voice-study scripts.
Copied verbatim from the stimulus-generation code, which is not part of the
release.
"""
from __future__ import annotations
import re
HIST_SPLIT = re.compile(r"role: (user|model)\\n response: ") # literal \n, as in the survey JS
def normalize(text: str) -> list[str]:
text = text.lower()
text = re.sub(r"[^a-z0-9' ]+", " ", text)
return text.split()
def wer(ref: list[str], hyp: list[str]) -> float:
"""Standard word error rate via edit distance."""
d = list(range(len(hyp) + 1))
for i, r in enumerate(ref, 1):
prev, d[0] = d[0], i
for j, h in enumerate(hyp, 1):
cur = min(d[j] + 1, d[j - 1] + 1, prev + (r != h))
prev, d[j] = d[j], cur
return d[-1] / max(1, len(ref))
def parse_history(prev_hist: str) -> list[tuple[str, str]]:
"""-> [(role, text), ...] in order; identical logic to the survey page."""
parts = [p for p in HIST_SPLIT.split(prev_hist)]
turns, i = [], 0
# split() yields [pre, role, text, role, text, ...]; pre is empty or noise
while i + 1 < len(parts):
if parts[i] in ("user", "model") and parts[i + 1].strip():
turns.append((parts[i], parts[i + 1].strip()))
i += 2
else:
i += 1
return turns
|