tzzht / Subator

Subator is a simple srt subtitle production tool.

What is Subator

Subator is a simple bilingual srt subtitle production tool. The tool comprises five core modules, implementing functions for downloading, transcribing, translating, segmenting, and aligning.

• Downloader: Utilizes the pytube package to download video and audio streams from YouTube.
• Transcriber: Employs openai Whisper for speech recognition on audio streams, providing word-level timestamps. Utilizes fullstop-deep-punctuation-prediction for sentence segmentation,
• Translator: Uses llm(gpt-3.5, qwen-max, glm-4) for translating sentences one by one.
• Spliter: Utilizes spaCy to segment sentences into short fragments, with Chinese not exceeding 33 characters and English not exceeding 80 letters.
• Aligner: Reads the word-level timestamps, aligns fragmens, and utilizes the srt package to output subtitles.

Prerequisites

• (optional) miniconda
  • https://docs.anaconda.com/free/miniconda/miniconda-install/
• pytubefix
  • pip install pytubefix
• Whisper
  • https://github.com/openai/whisper
• fullstop-deep-punctuation-prediction
  • https://github.com/oliverguhr/fullstop-deep-punctuation-prediction
• zhipuai
  • https://open.bigmodel.cn/dev/api
• dashscope
  • https://help.aliyun.com/zh/dashscope/developer-reference/
• spaCy
  • https://spacy.io/usage
  • conda install -c conda-forge spacy
  • python -m spacy download zh_core_web_trf
  • python -m spacy download en_core_web_trf
• srt
  • pip install -U srt
• webm
  • https://www.fnordware.com/WebM/
• opencc
  • pip install opencc
• openai
  • pip install openai

Getting Started

• Modify SAVE_DIR and API_KEY in src\subator_constants.py, python .\src\main.py --url "https://www.youtube.com/videoxxxxx"

• If you are using main.py, Subator will create a save_dir\video_author\video_title\resources folder in save_dir.

• The downloader will download video and audio streams to resources\video.webm and resources\audio.webm.

• The transcriber will transcribe resources\audio.webm and segment it with fullstop-deep-punctuation-prediction. Then output the results to resources\. Other modules will use two files: timestamps.json containing word-level timestamps, sentences.txt containing transcription sentences.

• The translator will read text from sentences.txt, and use the LLM interface for translation. Translation results will be saved in sentences_translated.txt, it has corresponding lines with sentences.txt. LLM outputs may sometimes deviate from expectations (containing context, additional explanatory statements, multiple lines, error due to safety checks, etc.). After translation, search for 'Please check the response.' in translator.log to locate potential errors and modify the corresponding lines in sentences_translated.txt. Do not modify sentences.txt as its content corresponds to word-level timestamps.

• The spliter will read sentences.txt and sentences_translated.txt, then split sentences into smaller fragments using simple strategies and spaCy. Results will be saved in fragments.json.

• The aligner will read fragments.json and timestamps.json to generate the final subtitle file, which will be saved in the save_dir\video_author\video_title folder.

• Subator can only help reduce the time it takes to create subtitles. The generated subtitles need to be proofread using tools like PR or SubtitleEdit before use.