54 lines
2.8 KiB
Python
54 lines
2.8 KiB
Python
|
"""Strict ordinary Whisper token-time sentence subdivision, never label driven."""
|
||
|
|
import re
|
||
|
|
|
||
|
|
POLICY = "ordinary-token-sentences-v1;punctuation-boundaries;exact-text;no-dtw;no-labels"
|
||
|
|
# Pinned multilingual Whisper vocabulary: ordinary text ids precede EOT.
|
||
|
|
TEXT_TOKEN_LIMIT = 50257
|
||
|
|
BOUNDARY = re.compile(r'[.!?。！？]["\u201d\u2019\')\]]*$')
|
||
|
|
|
||
|
|
|
||
|
|
def aligned_segments(document, offset_ms, duration_ms):
|
||
|
|
from run import segments
|
||
|
|
originals = segments(document, offset_ms, duration_ms)
|
||
|
|
result = []
|
||
|
|
raw_rows = [row for row in document["transcription"] if row["text"].strip()]
|
||
|
|
for index, (row, original) in enumerate(zip(raw_rows, originals)):
|
||
|
|
tokens = row.get("tokens")
|
||
|
|
if not isinstance(tokens, list) or not 0 < len(tokens) <= 4096:
|
||
|
|
raise ValueError("Missing or oversized token times")
|
||
|
|
accepted = []
|
||
|
|
previous = 0
|
||
|
|
for token in tokens:
|
||
|
|
token_id, text = token.get("id"), token.get("text")
|
||
|
|
if type(token_id) is not int or token_id < 0 or not isinstance(text, str):
|
||
|
|
raise ValueError("Invalid token identity")
|
||
|
|
# Structural tokens must never become recognized text candidates.
|
||
|
|
if token_id >= TEXT_TOKEN_LIMIT:
|
||
|
|
if not text.startswith("[_") or not text.endswith("]"):
|
||
|
|
raise ValueError("Unrecognized special token")
|
||
|
|
continue
|
||
|
|
times = token.get("offsets", {})
|
||
|
|
start, end = times.get("from"), times.get("to")
|
||
|
|
if (type(start) is not int or type(end) is not int
|
||
|
|
or not 0 <= start <= end <= duration_ms or start < previous
|
||
|
|
or not row["offsets"]["from"] <= start <= end <= row["offsets"]["to"]):
|
||
|
|
raise ValueError("Invalid ordinary token timestamps")
|
||
|
|
previous = start
|
||
|
|
accepted.append((text, start, end))
|
||
|
|
if "".join(token[0] for token in accepted) != row["text"]:
|
||
|
|
raise ValueError("Token text differs from original recognized segment")
|
||
|
|
phrase = []
|
||
|
|
for token_index, token in enumerate(accepted):
|
||
|
|
phrase.append(token)
|
||
|
|
if BOUNDARY.search(token[0]) or token_index == len(accepted) - 1:
|
||
|
|
text = "".join(part[0] for part in phrase)
|
||
|
|
if text.strip():
|
||
|
|
result.append({"start_ms": offset_ms + phrase[0][1],
|
||
|
|
"end_ms": offset_ms + phrase[-1][2], "text": text,
|
||
|
|
"original_segment_index": index,
|
||
|
|
"original_start_ms": original["start_ms"],
|
||
|
|
"original_end_ms": original["end_ms"]})
|
||
|
|
phrase = []
|
||
|
|
if len(result) > 10000:
|
||
|
|
raise ValueError("Oversized aligned passage pool")
|
||
|
|
return result
|