Chinese Transcription Tool

A Python tool that transcribes Chinese audio files and formats them into HSK exercise materials with pinyin and English translations.

Requirements

• Python 3.7+
• OpenAI API key
• uv package manager
• openai Python package
• just command runner (for batch processing)

Installation

1. Install the required package:

uv add openai

2. Set up your OpenAI API key:

export OPENAI_API_KEY="your-api-key-here"

Usage

Single File Transcription

Basic usage:

uv run index.py path/to/your/audio.mp3

With custom model:

uv run index.py path/to/your/audio.mp3 --model gpt-4o-mini

Batch Transcription

For processing multiple MP3 files in a folder, use the batch transcription script:

just transcribe <folder_path> [parallelism]

Examples:

# Process all MP3 files in ~/Desktop/hsk3 with 4 parallel jobs (default)
just transcribe ~/Desktop/hsk3

# Process with 8 parallel jobs for faster processing
just transcribe ~/Desktop/hsk3 8

The batch script will:

• Process all MP3 files in the specified folder
• Only process files with names greater than 8 (filtering logic)
• Create a transcribe/ directory with corresponding .txt output files
• Run multiple transcriptions in parallel for efficiency

Command Line Arguments

Single File (index.py)

• mp3_file_path (required): Path to the MP3 file to transcribe
• --model (optional): OpenAI model to use for processing (default: gpt-4o)

Batch Processing (just transcribe)

• folder_path (required): Path to folder containing MP3 files
• parallelism (optional): Number of parallel jobs (default: 4)

Output Format

The tool generates structured output in the following format:

== Part number: ... ==

# Exercise number: ..
## Question number: 
- Pinyin: ... (with proper tones and capitalization)
- Simplified chinese: ...
- Translation in english

Example

uv run index.py lesson1.mp3

Output:

== Part 1: Listening Comprehension ==

# Exercise 1: Daily Conversations
## Question 1: 
- Pinyin: Nǐ hǎo, nǐ jiào shén me míng zi?
- Simplified chinese: 你好，你叫什么名字？
- Translation in english: Hello, what is your name?

Additional tricks

How to get an audio trimmed until a certain blank:

ffmpeg -i 10.mp3 -af silencedetect=noise=-30dB:d=5 -f null - 2>&1 | grep silence_end | awk 'NR==6' | awk -F ' ' '{print int($5)-int(8)+1}' | xargs -I {} ffmpeg -i 10.mp3 -ss 00:00:00 -t {} output.mp3

Will stop just before the silence number 6 of the file 10.mp3

Generate one pdf per lesson

for f in *.txt; do
  paps --paper=a4 --font="Noto Sans CJK SC 11" "$f" | ps2pdf - "${f%.txt}.pdf"
done

License

This project is licensed under the MIT License - see the LICENSE file for details.