365 lines
13 KiB
Python
365 lines
13 KiB
Python
|
import json
|
|||
|
|
from core.utils import *
|
|||
|
|
|
|||
|
|
## ================================================================
|
|||
|
|
# @ step4_splitbymeaning.py
|
|||
|
|
def get_split_prompt(sentence, num_parts = 2, word_limit = 20):
|
|||
|
|
language = load_key("whisper.detected_language")
|
|||
|
|
split_prompt = f"""
|
|||
|
|
## Role
|
|||
|
|
You are a professional Netflix subtitle splitter in **{language}**.
|
|||
|
|
|
|||
|
|
## Task
|
|||
|
|
Split the given subtitle text into **{num_parts}** parts, each less than **{word_limit}** words.
|
|||
|
|
|
|||
|
|
1. Maintain sentence meaning coherence according to Netflix subtitle standards
|
|||
|
|
2. MOST IMPORTANT: Keep parts roughly equal in length (minimum 3 words each)
|
|||
|
|
3. Split at natural points like punctuation marks or conjunctions
|
|||
|
|
4. If provided text is repeated words, simply split at the middle of the repeated words.
|
|||
|
|
|
|||
|
|
## Steps
|
|||
|
|
1. Analyze the sentence structure, complexity, and key splitting challenges
|
|||
|
|
2. Generate two alternative splitting approaches with [br] tags at split positions
|
|||
|
|
3. Compare both approaches highlighting their strengths and weaknesses
|
|||
|
|
4. Choose the best splitting approach
|
|||
|
|
|
|||
|
|
## Given Text
|
|||
|
|
<split_this_sentence>
|
|||
|
|
{sentence}
|
|||
|
|
</split_this_sentence>
|
|||
|
|
|
|||
|
|
## Output in only JSON format and no other text
|
|||
|
|
```json
|
|||
|
|
{{
|
|||
|
|
"analysis": "Brief description of sentence structure, complexity, and key splitting challenges",
|
|||
|
|
"split1": "First splitting approach with [br] tags at split positions",
|
|||
|
|
"split2": "Alternative splitting approach with [br] tags at split positions",
|
|||
|
|
"assess": "Comparison of both approaches highlighting their strengths and weaknesses",
|
|||
|
|
"choice": "1 or 2"
|
|||
|
|
}}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
Note: Start you answer with ```json and end with ```, do not add any other text.
|
|||
|
|
""".strip()
|
|||
|
|
return split_prompt
|
|||
|
|
|
|||
|
|
"""{{
|
|||
|
|
"analysis": "Brief analysis of the text structure",
|
|||
|
|
"split": "Complete sentence with [br] tags at split positions"
|
|||
|
|
}}"""
|
|||
|
|
|
|||
|
|
## ================================================================
|
|||
|
|
# @ step4_1_summarize.py
|
|||
|
|
def get_summary_prompt(source_content, custom_terms_json=None):
|
|||
|
|
src_lang = load_key("whisper.detected_language")
|
|||
|
|
tgt_lang = load_key("target_language")
|
|||
|
|
|
|||
|
|
# add custom terms note
|
|||
|
|
terms_note = ""
|
|||
|
|
if custom_terms_json:
|
|||
|
|
terms_list = []
|
|||
|
|
for term in custom_terms_json['terms']:
|
|||
|
|
terms_list.append(f"- {term['src']}: {term['tgt']} ({term['note']})")
|
|||
|
|
terms_note = "\n### Existing Terms\nPlease exclude these terms in your extraction:\n" + "\n".join(terms_list)
|
|||
|
|
|
|||
|
|
summary_prompt = f"""
|
|||
|
|
## Role
|
|||
|
|
You are a video translation expert and terminology consultant, specializing in {src_lang} comprehension and {tgt_lang} expression optimization.
|
|||
|
|
|
|||
|
|
## Task
|
|||
|
|
For the provided {src_lang} video text:
|
|||
|
|
1. Summarize main topic in two sentences
|
|||
|
|
2. Extract professional terms/names with {tgt_lang} translations (excluding existing terms)
|
|||
|
|
3. Provide brief explanation for each term
|
|||
|
|
|
|||
|
|
{terms_note}
|
|||
|
|
|
|||
|
|
Steps:
|
|||
|
|
1. Topic Summary:
|
|||
|
|
- Quick scan for general understanding
|
|||
|
|
- Write two sentences: first for main topic, second for key point
|
|||
|
|
2. Term Extraction:
|
|||
|
|
- Mark professional terms and names (excluding those listed in Existing Terms)
|
|||
|
|
- Provide {tgt_lang} translation or keep original
|
|||
|
|
- Add brief explanation
|
|||
|
|
- Extract less than 15 terms
|
|||
|
|
|
|||
|
|
## INPUT
|
|||
|
|
<text>
|
|||
|
|
{source_content}
|
|||
|
|
</text>
|
|||
|
|
|
|||
|
|
## Output in only JSON format and no other text
|
|||
|
|
{{
|
|||
|
|
"theme": "Two-sentence video summary",
|
|||
|
|
"terms": [
|
|||
|
|
{{
|
|||
|
|
"src": "{src_lang} term",
|
|||
|
|
"tgt": "{tgt_lang} translation or original",
|
|||
|
|
"note": "Brief explanation"
|
|||
|
|
}},
|
|||
|
|
...
|
|||
|
|
]
|
|||
|
|
}}
|
|||
|
|
|
|||
|
|
## Example
|
|||
|
|
{{
|
|||
|
|
"theme": "本视频介绍人工智能在医疗领域的应用现状。重点展示了AI在医学影像诊断和药物研发中的突破性进展。",
|
|||
|
|
"terms": [
|
|||
|
|
{{
|
|||
|
|
"src": "Machine Learning",
|
|||
|
|
"tgt": "机器学习",
|
|||
|
|
"note": "AI的核心技术，通过数据训练实现智能决策"
|
|||
|
|
}},
|
|||
|
|
{{
|
|||
|
|
"src": "CNN",
|
|||
|
|
"tgt": "CNN",
|
|||
|
|
"note": "卷积神经网络，用于医学图像识别的深度学习模型"
|
|||
|
|
}}
|
|||
|
|
]
|
|||
|
|
}}
|
|||
|
|
|
|||
|
|
Note: Start you answer with ```json and end with ```, do not add any other text.
|
|||
|
|
""".strip()
|
|||
|
|
return summary_prompt
|
|||
|
|
|
|||
|
|
## ================================================================
|
|||
|
|
# @ step5_translate.py & translate_lines.py
|
|||
|
|
def generate_shared_prompt(previous_content_prompt, after_content_prompt, summary_prompt, things_to_note_prompt):
|
|||
|
|
return f'''### Context Information
|
|||
|
|
<previous_content>
|
|||
|
|
{previous_content_prompt}
|
|||
|
|
</previous_content>
|
|||
|
|
|
|||
|
|
<subsequent_content>
|
|||
|
|
{after_content_prompt}
|
|||
|
|
</subsequent_content>
|
|||
|
|
|
|||
|
|
### Content Summary
|
|||
|
|
{summary_prompt}
|
|||
|
|
|
|||
|
|
### Points to Note
|
|||
|
|
{things_to_note_prompt}'''
|
|||
|
|
|
|||
|
|
def get_prompt_faithfulness(lines, shared_prompt):
|
|||
|
|
TARGET_LANGUAGE = load_key("target_language")
|
|||
|
|
# Split lines by \n
|
|||
|
|
line_splits = lines.split('\n')
|
|||
|
|
|
|||
|
|
json_dict = {}
|
|||
|
|
for i, line in enumerate(line_splits, 1):
|
|||
|
|
json_dict[f"{i}"] = {"origin": line, "direct": f"direct {TARGET_LANGUAGE} translation {i}."}
|
|||
|
|
json_format = json.dumps(json_dict, indent=2, ensure_ascii=False)
|
|||
|
|
|
|||
|
|
src_language = load_key("whisper.detected_language")
|
|||
|
|
prompt_faithfulness = f'''
|
|||
|
|
## Role
|
|||
|
|
You are a professional Netflix subtitle translator, fluent in both {src_language} and {TARGET_LANGUAGE}, as well as their respective cultures.
|
|||
|
|
Your expertise lies in accurately understanding the semantics and structure of the original {src_language} text and faithfully translating it into {TARGET_LANGUAGE} while preserving the original meaning.
|
|||
|
|
|
|||
|
|
## Task
|
|||
|
|
We have a segment of original {src_language} subtitles that need to be directly translated into {TARGET_LANGUAGE}. These subtitles come from a specific context and may contain specific themes and terminology.
|
|||
|
|
|
|||
|
|
1. Translate the original {src_language} subtitles into {TARGET_LANGUAGE} line by line
|
|||
|
|
2. Ensure the translation is faithful to the original, accurately conveying the original meaning
|
|||
|
|
3. Consider the context and professional terminology
|
|||
|
|
|
|||
|
|
{shared_prompt}
|
|||
|
|
|
|||
|
|
<translation_principles>
|
|||
|
|
1. Faithful to the original: Accurately convey the content and meaning of the original text, without arbitrarily changing, adding, or omitting content.
|
|||
|
|
2. Accurate terminology: Use professional terms correctly and maintain consistency in terminology.
|
|||
|
|
3. Understand the context: Fully comprehend and reflect the background and contextual relationships of the text.
|
|||
|
|
</translation_principles>
|
|||
|
|
|
|||
|
|
## INPUT
|
|||
|
|
<subtitles>
|
|||
|
|
{lines}
|
|||
|
|
</subtitles>
|
|||
|
|
|
|||
|
|
## Output in only JSON format and no other text
|
|||
|
|
```json
|
|||
|
|
{json_format}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
Note: Start you answer with ```json and end with ```, do not add any other text.
|
|||
|
|
'''
|
|||
|
|
return prompt_faithfulness.strip()
|
|||
|
|
|
|||
|
|
|
|||
|
|
def get_prompt_expressiveness(faithfulness_result, lines, shared_prompt):
|
|||
|
|
TARGET_LANGUAGE = load_key("target_language")
|
|||
|
|
json_format = {
|
|||
|
|
key: {
|
|||
|
|
"origin": value["origin"],
|
|||
|
|
"direct": value["direct"],
|
|||
|
|
"reflect": "your reflection on direct translation",
|
|||
|
|
"free": "your free translation"
|
|||
|
|
}
|
|||
|
|
for key, value in faithfulness_result.items()
|
|||
|
|
}
|
|||
|
|
json_format = json.dumps(json_format, indent=2, ensure_ascii=False)
|
|||
|
|
|
|||
|
|
src_language = load_key("whisper.detected_language")
|
|||
|
|
prompt_expressiveness = f'''
|
|||
|
|
## Role
|
|||
|
|
You are a professional Netflix subtitle translator and language consultant.
|
|||
|
|
Your expertise lies not only in accurately understanding the original {src_language} but also in optimizing the {TARGET_LANGUAGE} translation to better suit the target language's expression habits and cultural background.
|
|||
|
|
|
|||
|
|
## Task
|
|||
|
|
We already have a direct translation version of the original {src_language} subtitles.
|
|||
|
|
Your task is to reflect on and improve these direct translations to create more natural and fluent {TARGET_LANGUAGE} subtitles.
|
|||
|
|
|
|||
|
|
1. Analyze the direct translation results line by line, pointing out existing issues
|
|||
|
|
2. Provide detailed modification suggestions
|
|||
|
|
3. Perform free translation based on your analysis
|
|||
|
|
4. Do not add comments or explanations in the translation, as the subtitles are for the audience to read
|
|||
|
|
5. Do not leave empty lines in the free translation, as the subtitles are for the audience to read
|
|||
|
|
|
|||
|
|
{shared_prompt}
|
|||
|
|
|
|||
|
|
<Translation Analysis Steps>
|
|||
|
|
Please use a two-step thinking process to handle the text line by line:
|
|||
|
|
|
|||
|
|
1. Direct Translation Reflection:
|
|||
|
|
- Evaluate language fluency
|
|||
|
|
- Check if the language style is consistent with the original text
|
|||
|
|
- Check the conciseness of the subtitles, point out where the translation is too wordy
|
|||
|
|
|
|||
|
|
2. {TARGET_LANGUAGE} Free Translation:
|
|||
|
|
- Aim for contextual smoothness and naturalness, conforming to {TARGET_LANGUAGE} expression habits
|
|||
|
|
- Ensure it's easy for {TARGET_LANGUAGE} audience to understand and accept
|
|||
|
|
- Adapt the language style to match the theme (e.g., use casual language for tutorials, professional terminology for technical content, formal language for documentaries)
|
|||
|
|
</Translation Analysis Steps>
|
|||
|
|
|
|||
|
|
## INPUT
|
|||
|
|
<subtitles>
|
|||
|
|
{lines}
|
|||
|
|
</subtitles>
|
|||
|
|
|
|||
|
|
## Output in only JSON format and no other text
|
|||
|
|
```json
|
|||
|
|
{json_format}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
Note: Start you answer with ```json and end with ```, do not add any other text.
|
|||
|
|
'''
|
|||
|
|
return prompt_expressiveness.strip()
|
|||
|
|
|
|||
|
|
|
|||
|
|
## ================================================================
|
|||
|
|
# @ step6_splitforsub.py
|
|||
|
|
def get_align_prompt(src_sub, tr_sub, src_part):
|
|||
|
|
targ_lang = load_key("target_language")
|
|||
|
|
src_lang = load_key("whisper.detected_language")
|
|||
|
|
src_splits = src_part.split('\n')
|
|||
|
|
num_parts = len(src_splits)
|
|||
|
|
src_part = src_part.replace('\n', ' [br] ')
|
|||
|
|
align_parts_json = ','.join(
|
|||
|
|
f'''
|
|||
|
|
{{
|
|||
|
|
"src_part_{i+1}": "{src_splits[i]}",
|
|||
|
|
"target_part_{i+1}": "Corresponding aligned {targ_lang} subtitle part"
|
|||
|
|
}}''' for i in range(num_parts)
|
|||
|
|
)
|
|||
|
|
|
|||
|
|
align_prompt = f'''
|
|||
|
|
## Role
|
|||
|
|
You are a Netflix subtitle alignment expert fluent in both {src_lang} and {targ_lang}.
|
|||
|
|
|
|||
|
|
## Task
|
|||
|
|
We have {src_lang} and {targ_lang} original subtitles for a Netflix program, as well as a pre-processed split version of {src_lang} subtitles.
|
|||
|
|
Your task is to create the best splitting scheme for the {targ_lang} subtitles based on this information.
|
|||
|
|
|
|||
|
|
1. Analyze the word order and structural correspondence between {src_lang} and {targ_lang} subtitles
|
|||
|
|
2. Split the {targ_lang} subtitles according to the pre-processed {src_lang} split version
|
|||
|
|
3. Never leave empty lines. If it's difficult to split based on meaning, you may appropriately rewrite the sentences that need to be aligned
|
|||
|
|
4. Do not add comments or explanations in the translation, as the subtitles are for the audience to read
|
|||
|
|
|
|||
|
|
## INPUT
|
|||
|
|
<subtitles>
|
|||
|
|
{src_lang} Original: "{src_sub}"
|
|||
|
|
{targ_lang} Original: "{tr_sub}"
|
|||
|
|
Pre-processed {src_lang} Subtitles ([br] indicates split points): {src_part}
|
|||
|
|
</subtitles>
|
|||
|
|
|
|||
|
|
## Output in only JSON format and no other text
|
|||
|
|
```json
|
|||
|
|
{{
|
|||
|
|
"analysis": "Brief analysis of word order, structure, and semantic correspondence between two subtitles",
|
|||
|
|
"align": [
|
|||
|
|
{align_parts_json}
|
|||
|
|
]
|
|||
|
|
}}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
Note: Start you answer with ```json and end with ```, do not add any other text.
|
|||
|
|
'''.strip()
|
|||
|
|
return align_prompt
|
|||
|
|
|
|||
|
|
## ================================================================
|
|||
|
|
# @ step8_gen_audio_task.py @ step10_gen_audio.py
|
|||
|
|
def get_subtitle_trim_prompt(text, duration):
|
|||
|
|
|
|||
|
|
rule = '''Consider a. Reducing filler words without modifying meaningful content. b. Omitting unnecessary modifiers or pronouns, for example:
|
|||
|
|
- "Please explain your thought process" can be shortened to "Please explain thought process"
|
|||
|
|
- "We need to carefully analyze this complex problem" can be shortened to "We need to analyze this problem"
|
|||
|
|
- "Let's discuss the various different perspectives on this topic" can be shortened to "Let's discuss different perspectives on this topic"
|
|||
|
|
- "Can you describe in detail your experience from yesterday" can be shortened to "Can you describe yesterday's experience" '''
|
|||
|
|
|
|||
|
|
trim_prompt = f'''
|
|||
|
|
## Role
|
|||
|
|
You are a professional subtitle editor, editing and optimizing lengthy subtitles that exceed voiceover time before handing them to voice actors.
|
|||
|
|
Your expertise lies in cleverly shortening subtitles slightly while ensuring the original meaning and structure remain unchanged.
|
|||
|
|
|
|||
|
|
## INPUT
|
|||
|
|
<subtitles>
|
|||
|
|
Subtitle: "{text}"
|
|||
|
|
Duration: {duration} seconds
|
|||
|
|
</subtitles>
|
|||
|
|
|
|||
|
|
## Processing Rules
|
|||
|
|
{rule}
|
|||
|
|
|
|||
|
|
## Processing Steps
|
|||
|
|
Please follow these steps and provide the results in the JSON output:
|
|||
|
|
1. Analysis: Briefly analyze the subtitle's structure, key information, and filler words that can be omitted.
|
|||
|
|
2. Trimming: Based on the rules and analysis, optimize the subtitle by making it more concise according to the processing rules.
|
|||
|
|
|
|||
|
|
## Output in only JSON format and no other text
|
|||
|
|
```json
|
|||
|
|
{{
|
|||
|
|
"analysis": "Brief analysis of the subtitle, including structure, key information, and potential processing locations",
|
|||
|
|
"result": "Optimized and shortened subtitle in the original subtitle language"
|
|||
|
|
}}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
Note: Start you answer with ```json and end with ```, do not add any other text.
|
|||
|
|
'''.strip()
|
|||
|
|
return trim_prompt
|
|||
|
|
|
|||
|
|
## ================================================================
|
|||
|
|
# @ tts_main
|
|||
|
|
def get_correct_text_prompt(text):
|
|||
|
|
return f'''
|
|||
|
|
## Role
|
|||
|
|
You are a text cleaning expert for TTS (Text-to-Speech) systems.
|
|||
|
|
|
|||
|
|
## Task
|
|||
|
|
Clean the given text by:
|
|||
|
|
1. Keep only basic punctuation (.,?!)
|
|||
|
|
2. Preserve the original meaning
|
|||
|
|
|
|||
|
|
## INPUT
|
|||
|
|
{text}
|
|||
|
|
|
|||
|
|
## Output in only JSON format and no other text
|
|||
|
|
```json
|
|||
|
|
{{
|
|||
|
|
"text": "cleaned text here"
|
|||
|
|
}}
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
Note: Start you answer with ```json and end with ```, do not add any other text.
|
|||
|
|
'''.strip()
|