Files
speaking-practice/backend/subtitle_ai.py
T

122 lines
4.5 KiB
Python

from __future__ import annotations
from .subtitle_tools import SubtitleCue, build_vtt, clean_english_subtitle
TRANSLATION_SYSTEM_PROMPT = """
你是英语学习产品的字幕处理助手。输入是按顺序排列的英文字幕。
请只返回 JSON,格式为 {"translations":[{"index":1,"zh":"中文翻译"}]}。
要求:
1. 只翻译英文文本,不要改写原文含义。
2. 不要修改时间轴,不要生成任何 start/end/time 字段。
3. 不要增删字幕条数,每个输入 index 必须返回且只返回一条中文。
4. translations 可以乱序,但 index 必须和输入完全一致。
5. 中文自然,适合英语学习;不要解释。
""".strip()
def process_subtitles_with_ai(
english_content: str,
client,
batch_size: int = 0,
) -> dict[str, object]:
cleaned = clean_english_subtitle(english_content)
if not cleaned.cues:
raise ValueError("英文字幕为空,无法处理。")
english_cues = cleaned.cues
translations = translate_cues(english_cues, client, batch_size)
zh_cues = []
for index, cue in enumerate(english_cues, 1):
translation = translations[index]
zh_text = str(translation.get("zh", "")).strip()
if not zh_text:
raise ValueError(f"AI 返回了空翻译:index {index}")
zh_cues.append({"start": cue.start, "end": cue.end, "text": zh_text})
translation_batches = count_batches(len(english_cues), batch_size)
return {
"subtitle_en_content": build_vtt(english_cues),
"subtitle_zh_content": build_vtt(zh_cues),
"report": {
"original_count": cleaned.original_count,
"cleaned_count": cleaned.cleaned_count,
"duplicate_count": cleaned.duplicate_count,
"empty_count": cleaned.empty_count,
"translation_count": len(zh_cues),
"translation_batches": translation_batches,
},
}
def translate_cues(
cues: list[SubtitleCue],
client,
batch_size: int,
) -> dict[int, dict[str, object]]:
translations: dict[int, dict[str, object]] = {}
indexed_cues = list(enumerate(cues, 1))
effective_batch_size = int(batch_size or 0)
batches = (
[indexed_cues]
if effective_batch_size <= 0 or effective_batch_size >= len(indexed_cues)
else [indexed_cues[start:start + effective_batch_size] for start in range(0, len(indexed_cues), effective_batch_size)]
)
for batch in batches:
expected_indexes = {index for index, _cue in batch}
collect_translations(batch, expected_indexes, translations, client)
missing = sorted(expected_indexes - translations.keys())
if missing:
missing_batch = [(index, cue) for index, cue in batch if index in missing]
collect_translations(missing_batch, set(missing), translations, client)
missing = sorted(expected_indexes - translations.keys())
if missing:
for index, cue in [(index, cue) for index, cue in batch if index in missing]:
collect_translations([(index, cue)], {index}, translations, client)
missing = sorted(expected_indexes - translations.keys())
if missing:
raise ValueError(f"AI 返回的翻译条数和英文字幕条数不一致,缺少 index:{missing[:10]}")
return translations
def collect_translations(
batch: list[tuple[int, SubtitleCue]],
expected_indexes: set[int],
translations: dict[int, dict[str, object]],
client,
) -> None:
if not batch:
return
payload = {
"cues": [
{"index": index, "text": cue.text}
for index, cue in batch
]
}
response = client.chat_json(TRANSLATION_SYSTEM_PROMPT, payload)
batch_translations = response.get("translations")
if not isinstance(batch_translations, list):
raise ValueError("AI 返回结果缺少 translations。")
for item in batch_translations:
if not isinstance(item, dict):
continue
try:
item_index = int(item.get("index"))
except (TypeError, ValueError):
continue
if item_index in expected_indexes:
translations[item_index] = item
def count_batches(cue_count: int, batch_size: int) -> int:
effective_batch_size = int(batch_size or 0)
if cue_count <= 0:
return 0
if effective_batch_size <= 0 or effective_batch_size >= cue_count:
return 1
return (cue_count + effective_batch_size - 1) // effective_batch_size