mirror of
https://github.com/unshackle-dl/unshackle.git
synced 2026-11-15 21:45:10 +00:00
fix(subs): stop segmented WebVTT merge from stripping formatting
Merge cues at the string level instead of round-tripping through pycaption, which dropped inline tags and cue settings. STYLE and REGION blocks are kept, duplicate boundary cues splice correctly, and a bad timing line now skips the cue instead of crashing. Also rank pysubs2 above SubtitleEdit for ASS to TTML/VTT since SubtitleEdit flattens inline styling there. Documented the measured backend fidelity table and added regression tests for the merge.
This commit is contained in:
@@ -154,6 +154,11 @@ class SubtitleEditBackend:
|
||||
return source in self.reads and target in self.writes
|
||||
|
||||
def rank(self, source: Codec, target: Codec) -> int:
|
||||
if source in (Codec.SubStationAlpha, Codec.SubStationAlphav4) and target in (
|
||||
Codec.TimedTextMarkupLang,
|
||||
Codec.WebVTT,
|
||||
):
|
||||
return 3
|
||||
return 0
|
||||
|
||||
def convert(self, source: Codec, src: Path, target: Codec, out: Path) -> None:
|
||||
|
||||
+57
-208
@@ -1,225 +1,74 @@
|
||||
import re
|
||||
import sys
|
||||
import typing
|
||||
from typing import Optional
|
||||
|
||||
import pysubs2
|
||||
from pycaption import Caption, CaptionList, CaptionNode, CaptionReadError, WebVTTReader, WebVTTWriter
|
||||
|
||||
from unshackle.core.config import config
|
||||
# Timing line: optional hours, verbatim settings tail.
|
||||
_TIMING = re.compile(r"^((?:\d+:)?\d{2}:\d{2}\.\d{3})[ \t]+-->[ \t]+((?:\d+:)?\d{2}:\d{2}\.\d{3})(.*)$")
|
||||
|
||||
|
||||
class CaptionListExt(CaptionList):
|
||||
@typing.no_type_check
|
||||
def __init__(self, iterable=None, layout_info=None):
|
||||
self.first_segment_mpegts = 0
|
||||
super().__init__(iterable, layout_info)
|
||||
|
||||
|
||||
class CaptionExt(Caption):
|
||||
@typing.no_type_check
|
||||
def __init__(self, start, end, nodes, style=None, layout_info=None, segment_index=0, mpegts=0, cue_time=0.0):
|
||||
style = style or {}
|
||||
self.segment_index: int = segment_index
|
||||
self.mpegts: float = mpegts
|
||||
self.cue_time: float = cue_time
|
||||
super().__init__(start, end, nodes, style, layout_info)
|
||||
|
||||
|
||||
class WebVTTReaderExt(WebVTTReader):
|
||||
# HLS extension support <https://datatracker.ietf.org/doc/html/rfc8216#section-3.5>
|
||||
RE_TIMESTAMP_MAP = re.compile(r"X-TIMESTAMP-MAP.*")
|
||||
RE_MPEGTS = re.compile(r"MPEGTS:(\d+)")
|
||||
RE_LOCAL = re.compile(r"LOCAL:((?:(\d{1,}):)?(\d{2}):(\d{2})\.(\d{3}))")
|
||||
|
||||
def _parse(self, lines: list[str]) -> CaptionList:
|
||||
captions = CaptionListExt()
|
||||
start = None
|
||||
end = None
|
||||
nodes: list[CaptionNode] = []
|
||||
layout_info = None
|
||||
found_timing = False
|
||||
segment_index = -1
|
||||
mpegts = 0
|
||||
cue_time = 0.0
|
||||
|
||||
# The first segment MPEGTS is needed to calculate the rest. It is possible that
|
||||
# the first segment contains no cue and is ignored by pycaption, this acts as a fallback.
|
||||
captions.first_segment_mpegts = 0
|
||||
|
||||
for i, line in enumerate(lines):
|
||||
if "-->" in line:
|
||||
found_timing = True
|
||||
timing_line = i
|
||||
last_start_time = captions[-1].start if captions else 0
|
||||
try:
|
||||
start, end, layout_info = self._parse_timing_line(line, last_start_time)
|
||||
except CaptionReadError as e:
|
||||
new_msg = f"{e.args[0]} (line {timing_line})"
|
||||
tb = sys.exc_info()[2]
|
||||
raise type(e)(new_msg).with_traceback(tb) from None
|
||||
|
||||
elif "" == line:
|
||||
if found_timing and nodes:
|
||||
found_timing = False
|
||||
caption = CaptionExt(
|
||||
start,
|
||||
end,
|
||||
nodes,
|
||||
layout_info=layout_info,
|
||||
segment_index=segment_index,
|
||||
mpegts=mpegts,
|
||||
cue_time=cue_time,
|
||||
)
|
||||
captions.append(caption)
|
||||
nodes = []
|
||||
|
||||
elif "WEBVTT" in line:
|
||||
# Merged segmented VTT doesn't have index information, track manually.
|
||||
segment_index += 1
|
||||
mpegts = 0
|
||||
cue_time = 0.0
|
||||
elif m := self.RE_TIMESTAMP_MAP.match(line):
|
||||
if r := self.RE_MPEGTS.search(m.group()):
|
||||
mpegts = int(r.group(1))
|
||||
|
||||
cue_time = self._parse_local(m.group())
|
||||
|
||||
# Early assignment in case the first segment contains no cue.
|
||||
if segment_index == 0:
|
||||
captions.first_segment_mpegts = mpegts
|
||||
|
||||
else:
|
||||
if found_timing:
|
||||
if nodes:
|
||||
nodes.append(CaptionNode.create_break())
|
||||
nodes.append(CaptionNode.create_text(self._decode(line)))
|
||||
else:
|
||||
# it's a comment or some metadata; ignore it
|
||||
pass
|
||||
|
||||
# Add a last caption if there are remaining nodes
|
||||
if nodes:
|
||||
caption = CaptionExt(start, end, nodes, layout_info=layout_info, segment_index=segment_index, mpegts=mpegts)
|
||||
captions.append(caption)
|
||||
|
||||
return captions
|
||||
|
||||
@staticmethod
|
||||
def _parse_local(string: str) -> float:
|
||||
"""
|
||||
Parse WebVTT LOCAL time and convert it to seconds.
|
||||
"""
|
||||
m = WebVTTReaderExt.RE_LOCAL.search(string)
|
||||
if not m:
|
||||
return 0
|
||||
|
||||
parsed = m.groups()
|
||||
if not parsed:
|
||||
return 0
|
||||
hours = int(parsed[1])
|
||||
minutes = int(parsed[2])
|
||||
seconds = int(parsed[3])
|
||||
milliseconds = int(parsed[4])
|
||||
return (milliseconds / 1000) + seconds + (minutes * 60) + (hours * 3600)
|
||||
def _timestamp_ms(ts: str) -> int:
|
||||
hms, ms = ts.rsplit(".", 1)
|
||||
parts = [int(p) for p in hms.split(":")]
|
||||
while len(parts) < 3:
|
||||
parts.insert(0, 0)
|
||||
hours, minutes, seconds = parts
|
||||
return ((hours * 60 + minutes) * 60 + seconds) * 1000 + int(ms)
|
||||
|
||||
|
||||
def merge_segmented_webvtt(vtt_raw: str, segment_durations: Optional[list[int]] = None, timescale: int = 1) -> str:
|
||||
"""
|
||||
Merge Segmented WebVTT data.
|
||||
Merge concatenated Segmented WebVTT documents into a single document.
|
||||
|
||||
Parameters:
|
||||
vtt_raw: The concatenated WebVTT files to merge. All WebVTT headers must be
|
||||
appropriately spaced apart, or it may produce unwanted effects like
|
||||
considering headers as captions, timestamp lines, etc.
|
||||
segment_durations: A list of each segment's duration. If not provided it will try
|
||||
to get it from the X-TIMESTAMP-MAP headers, specifically the MPEGTS number.
|
||||
timescale: The number of time units per second.
|
||||
Strips the per-segment WEBVTT and X-TIMESTAMP-MAP headers and splices cues
|
||||
repeated across a segment boundary (identical payload, <=1ms gap or overlap).
|
||||
Timing lines, cue settings and payload are kept verbatim, so inline formatting
|
||||
(span tags, entities, ASS-style overrides) survives untouched.
|
||||
|
||||
This parses the X-TIMESTAMP-MAP data to compute new absolute timestamps, replacing
|
||||
the old start and end timestamp values. All X-TIMESTAMP-MAP header information will
|
||||
be removed from the output as they are no longer of concern. Consider this function
|
||||
the opposite of a WebVTT Segmenter, a WebVTT Joiner of sorts.
|
||||
No timestamp offset is applied: segments are expected to carry absolute cue
|
||||
times already (same guard as shaka-player and N_m3u8DL-RE; re-offsetting
|
||||
double-shifts such streams). segment_durations/timescale are accepted for
|
||||
call-site compatibility and only matter if offsetting is ever reintroduced.
|
||||
|
||||
Algorithm borrowed from N_m3u8DL-RE and shaka-player.
|
||||
Unique STYLE/REGION blocks are kept (players that don't support them ignore
|
||||
them); cue identifiers and NOTE blocks are dropped; cues with a malformed
|
||||
timing line are skipped (downstream sanitizers catch worse).
|
||||
"""
|
||||
MPEG_TIMESCALE = 90_000
|
||||
# each kept cue: [start, end, settings, payload, normalized payload for dedupe]
|
||||
cues: list[list[str]] = []
|
||||
headers: list[str] = [] # unique STYLE/REGION blocks, first-seen order
|
||||
prev: Optional[list[str]] = None
|
||||
prev_end_ms = 0
|
||||
|
||||
# Check config for conversion method preference
|
||||
conversion_method = config.subtitle.get("conversion_method", "auto")
|
||||
use_pysubs2 = conversion_method in ("pysubs2", "auto")
|
||||
for block in re.split(r"\n[ \t]*\n", vtt_raw.replace("\r\n", "\n").replace("\r", "\n")):
|
||||
lines = [
|
||||
line
|
||||
for line in block.split("\n")
|
||||
if not (line.startswith("WEBVTT") or line.lstrip().startswith("X-TIMESTAMP-MAP"))
|
||||
]
|
||||
timing_i = next((i for i, line in enumerate(lines) if "-->" in line), None)
|
||||
if timing_i is None: # NOTE or other metadata block
|
||||
first = next((line.strip() for line in lines if line.strip()), "")
|
||||
if first.startswith(("STYLE", "REGION")):
|
||||
header = "\n".join(lines).strip("\n")
|
||||
if header not in headers: # segments repeat the same block
|
||||
headers.append(header)
|
||||
continue
|
||||
timing = _TIMING.match(lines[timing_i].strip())
|
||||
if not timing:
|
||||
continue
|
||||
start, end, settings = timing.groups()
|
||||
payload = [line for line in lines[timing_i + 1 :] if line.strip()]
|
||||
if not payload:
|
||||
continue
|
||||
|
||||
if use_pysubs2:
|
||||
# Try using pysubs2 first for more lenient parsing
|
||||
try:
|
||||
# Use pysubs2 to parse and normalize the VTT
|
||||
subs = pysubs2.SSAFile.from_string(vtt_raw)
|
||||
# Convert back to WebVTT string for pycaption processing
|
||||
normalized_vtt = subs.to_string("vtt")
|
||||
vtt = WebVTTReaderExt().read(normalized_vtt)
|
||||
except Exception:
|
||||
# Fall back to direct pycaption parsing
|
||||
vtt = WebVTTReaderExt().read(vtt_raw)
|
||||
else:
|
||||
# Use pycaption directly
|
||||
vtt = WebVTTReaderExt().read(vtt_raw)
|
||||
for lang in vtt.get_languages():
|
||||
prev_caption = None
|
||||
duplicate_index: list[int] = []
|
||||
captions = vtt.get_captions(lang)
|
||||
normalized = "\n".join(line.strip() for line in payload)
|
||||
if prev is not None and _timestamp_ms(start) - prev_end_ms <= 1 and normalized == prev[4]:
|
||||
prev[1] = end # splice: extend the kept cue, drop the duplicate
|
||||
prev_end_ms = _timestamp_ms(end)
|
||||
continue
|
||||
|
||||
# Some providers can produce "segment_index" values that are
|
||||
# outside the provided segment_durations list after normalization/merge.
|
||||
# This used to crash with IndexError and abort the entire download.
|
||||
if segment_durations and captions:
|
||||
max_idx = max(getattr(c, "segment_index", 0) for c in captions)
|
||||
if max_idx >= len(segment_durations):
|
||||
# Pad with the last known duration (or 0 if empty) so indexing is safe.
|
||||
pad_val = segment_durations[-1] if segment_durations else 0
|
||||
segment_durations = segment_durations + [pad_val] * (max_idx - len(segment_durations) + 1)
|
||||
prev = [start, end, settings, "\n".join(payload), normalized]
|
||||
prev_end_ms = _timestamp_ms(end)
|
||||
cues.append(prev)
|
||||
|
||||
if captions[0].segment_index == 0:
|
||||
first_segment_mpegts = captions[0].mpegts
|
||||
else:
|
||||
first_segment_mpegts = segment_durations[0] if segment_durations else captions.first_segment_mpegts
|
||||
|
||||
caption: CaptionExt
|
||||
for i, caption in enumerate(captions):
|
||||
# DASH WebVTT doesn't have MPEGTS timestamp like HLS. Instead,
|
||||
# calculate the timestamp from SegmentTemplate/SegmentList duration.
|
||||
likely_dash = first_segment_mpegts == 0 and caption.mpegts == 0
|
||||
if likely_dash and segment_durations:
|
||||
# Defensive: segment_index can still be out of range if captions are malformed.
|
||||
if caption.segment_index < 0 or caption.segment_index >= len(segment_durations):
|
||||
continue
|
||||
duration = segment_durations[caption.segment_index]
|
||||
caption.mpegts = MPEG_TIMESCALE * (duration / timescale)
|
||||
|
||||
if caption.mpegts == 0:
|
||||
continue
|
||||
|
||||
# Commeted to fix DSNP subs being out of sync and mistimed.
|
||||
# seconds = (caption.mpegts - first_segment_mpegts) / MPEG_TIMESCALE - caption.cue_time
|
||||
# offset = seconds * 1_000_000 # pycaption use microseconds
|
||||
|
||||
# if caption.start < offset:
|
||||
# caption.start += offset
|
||||
# caption.end += offset
|
||||
|
||||
# If the difference between current and previous captions is <=1ms
|
||||
# and the payload is equal then splice.
|
||||
if (
|
||||
prev_caption
|
||||
and not caption.is_empty()
|
||||
and (caption.start - prev_caption.end) <= 1000 # 1ms in microseconds
|
||||
and caption.get_text() == prev_caption.get_text()
|
||||
):
|
||||
prev_caption.end = caption.end
|
||||
duplicate_index.append(i)
|
||||
|
||||
prev_caption = caption
|
||||
|
||||
# Remove duplicate
|
||||
captions[:] = [c for c_index, c in enumerate(captions) if c_index not in set(duplicate_index)]
|
||||
|
||||
return WebVTTWriter().write(vtt)
|
||||
blocks = headers + [f"{c[0]} --> {c[1]}{c[2]}\n{c[3]}" for c in cues]
|
||||
return "WEBVTT\n\n" + "\n\n".join(blocks) + "\n"
|
||||
|
||||
Reference in New Issue
Block a user