125 lines
4.3 KiB
Python
125 lines
4.3 KiB
Python
from __future__ import annotations
|
||
|
||
from dataclasses import dataclass
|
||
from pathlib import Path
|
||
import subprocess
|
||
import tempfile
|
||
from typing import Callable
|
||
|
||
from .subtitle_tools import SubtitleCue, parse_timestamp
|
||
|
||
|
||
FIRST_SENTENCE_ALIGNMENT_PROMPT = """
|
||
你是字幕首句时间校准助手。你只需要分析这段音频的前 10 秒。
|
||
请找出给定英文字幕第一句在音频中真正开始说话的时间,单位是秒。
|
||
只返回 JSON:{"first_sentence_start":1.23,"matched_text":"...","confidence":0.95}
|
||
不要返回解释,不要校准后续句子。
|
||
""".strip()
|
||
|
||
|
||
@dataclass
|
||
class AlignmentResult:
|
||
applied: bool
|
||
offset_seconds: float
|
||
first_sentence_start: float | None
|
||
confidence: float
|
||
note: str
|
||
|
||
|
||
def align_first_sentence(video_path: Path, first_cue: SubtitleCue, client) -> AlignmentResult:
|
||
if not video_path.exists():
|
||
return AlignmentResult(False, 0.0, None, 0.0, "视频文件不存在,已跳过自动首句校准。")
|
||
|
||
with tempfile.TemporaryDirectory() as temp_dir:
|
||
audio_path = Path(temp_dir) / "first10.wav"
|
||
try:
|
||
extract_first_audio(video_path, audio_path)
|
||
response = client.audio_json(
|
||
FIRST_SENTENCE_ALIGNMENT_PROMPT,
|
||
audio_path,
|
||
f"第一句英文字幕:{first_cue.text}\n请只判断这句开始说话的秒数。",
|
||
)
|
||
except Exception as error:
|
||
return AlignmentResult(False, 0.0, None, 0.0, f"首句自动校准失败,已跳过:{error}")
|
||
|
||
return alignment_from_start(first_cue, response.get("first_sentence_start"), response.get("confidence"))
|
||
|
||
|
||
def align_first_sentence_with_asr(
|
||
video_path: Path,
|
||
first_cue: SubtitleCue,
|
||
asr_client,
|
||
extract_audio: Callable[[Path, Path], None] = None,
|
||
) -> AlignmentResult:
|
||
if not video_path.exists():
|
||
return AlignmentResult(False, 0.0, None, 0.0, "视频文件不存在,已跳过 OpenAI ASR 首句校准。")
|
||
|
||
audio_extractor = extract_audio or extract_first_audio
|
||
with tempfile.TemporaryDirectory() as temp_dir:
|
||
audio_path = Path(temp_dir) / "first10.wav"
|
||
try:
|
||
audio_extractor(video_path, audio_path)
|
||
response = asr_client.transcribe(audio_path)
|
||
except Exception as error:
|
||
return AlignmentResult(False, 0.0, None, 0.0, f"OpenAI ASR 首句校准失败,已跳过:{error}")
|
||
|
||
first_start = first_segment_start(response)
|
||
return alignment_from_start(first_cue, first_start, 0.9, source_name="OpenAI ASR")
|
||
|
||
|
||
def alignment_from_start(
|
||
first_cue: SubtitleCue,
|
||
start_value,
|
||
confidence_value=None,
|
||
source_name: str = "AI",
|
||
) -> AlignmentResult:
|
||
first_sentence_start = parse_alignment_second(start_value)
|
||
confidence = parse_alignment_second(confidence_value)
|
||
if first_sentence_start is None:
|
||
return AlignmentResult(False, 0.0, None, confidence or 0.0, f"{source_name} 没有返回有效的首句开始时间,已跳过自动校准。")
|
||
|
||
offset = first_sentence_start - parse_timestamp(first_cue.start)
|
||
if abs(offset) > 5:
|
||
return AlignmentResult(False, 0.0, first_sentence_start, confidence or 0.0, f"检测到偏移 {offset:.2f} 秒,超过 5 秒安全阈值,已跳过自动校准。")
|
||
|
||
return AlignmentResult(True, offset, first_sentence_start, confidence or 0.0, f"{source_name} 首句校准成功,整体平移 {offset:.2f} 秒。")
|
||
|
||
|
||
def first_segment_start(response: dict[str, object]) -> float | None:
|
||
segments = response.get("segments")
|
||
if isinstance(segments, list):
|
||
for segment in segments:
|
||
if isinstance(segment, dict):
|
||
start = parse_alignment_second(segment.get("start"))
|
||
text = str(segment.get("text", "")).strip()
|
||
if start is not None and text:
|
||
return start
|
||
return None
|
||
|
||
|
||
def extract_first_audio(video_path: Path, audio_path: Path) -> None:
|
||
command = [
|
||
"ffmpeg",
|
||
"-y",
|
||
"-i",
|
||
str(video_path),
|
||
"-ss",
|
||
"0",
|
||
"-t",
|
||
"10",
|
||
"-vn",
|
||
"-ac",
|
||
"1",
|
||
"-ar",
|
||
"16000",
|
||
str(audio_path),
|
||
]
|
||
subprocess.run(command, check=True, capture_output=True)
|
||
|
||
|
||
def parse_alignment_second(value) -> float | None:
|
||
try:
|
||
return float(value)
|
||
except (TypeError, ValueError):
|
||
return None
|