NLP-midterm

Vietnamese text processing and audio alignment project.

Features

• Sentence Splitting: Split Vietnamese text files into one sentence per line
• Sentence Counting: Count sentences in Vietnamese text files
• Audio Alignment: Align Vietnamese audio with text and cut into sentence segments

Installation

pip install -r requirements.txt

Note: For audio processing, you also need ffmpeg:

• Windows: Download from ffmpeg.org or use choco install ffmpeg
• macOS: brew install ffmpeg
• Linux: sudo apt-get install ffmpeg (Ubuntu/Debian)

Usage

Sentence Splitting

Split a text file into sentences (one per line):

python src/sentence_splitter.py <input_file> [-o output_file]

Example:

python src/sentence_splitter.py "data/Thiên thần nhỏ của tôi - Nguyễn Nhật Ánh.txt"

Sentence Counting

Count sentences in the data files:

cd src
python -c "from main import count_sentences_in_data_files; count_sentences_in_data_files()"

Audio Alignment

Option 1: WhisperX (Recommended for Better Accuracy)

WhisperX provides more accurate word-level timestamps than basic Whisper:

python src/align_vietnamese_audio_whisperx.py <audio_file> <sentences_file> [output_dir] [model] [device]

Example:

python src/align_vietnamese_audio_whisperx.py audio.wav "data/Text-ThienThanNhoCuaToi/Track 1.txt" output_audio base cpu

Parameters:

• audio_file: Path to audio file (wav, mp3, etc.)
• sentences_file: Path to processed text file (one sentence per line)
• output_dir: Output directory for audio segments (default: audio_segments)
• model: Whisper model size - tiny, base, small, medium, large (default: base)
• device: cpu or cuda (default: cpu)

Option 2: Basic Whisper (Faster but Less Accurate)

python src/align_vietnamese_audio.py <audio_file> <sentences_file> [output_dir] [model]

Output (both methods):

• sentence_00001.wav, sentence_00002.wav, ... - Audio segments
• sentence_00001.txt, sentence_00002.txt, ... - Corresponding text files
• timestamps.txt - Timestamps for each sentence
• transcription.txt - Full transcription

Option 3: Montreal Forced Aligner (MFA) - Most Accurate

For the highest accuracy, consider using MFA (Montreal Forced Aligner):

1. Install MFA: conda install -c conda-forge montreal-forced-alignment
2. Download Vietnamese acoustic model: mfa model download acoustic vietnamese_mfa
3. Use MFA command-line tool for alignment

MFA provides the most accurate forced alignment but requires more setup.

Project Structure

NLP-midterm/
├── data/                          # Text files
│   ├── *.txt                      # Original text files
│   └── *.processed.txt            # Processed (one sentence per line)
├── src/
│   ├── sentence_splitter.py              # Sentence splitting script
│   ├── main.py                           # Main processing and counting
│   ├── align_vietnamese_audio.py         # Basic Whisper alignment (faster)
│   └── align_vietnamese_audio_whisperx.py # WhisperX alignment (more accurate)
└── requirements.txt               # Python dependencies

Notes

• All text files use UTF-8 encoding to support Vietnamese diacritics
• For best accuracy: Use WhisperX (align_vietnamese_audio_whisperx.py) - provides better word-level timestamps
• For fastest processing: Use basic Whisper (align_vietnamese_audio.py)
• For highest accuracy: Consider MFA (Montreal Forced Aligner) - requires separate installation
• Recommended Whisper models: base or small for Vietnamese
• Audio loading uses librosa (no ffmpeg required for MP3 files)