whisperx-nemo-pipeline 1.0.4

Production-ready transcription and diarization pipeline with parallel processing

WhisperX-NeMo Pipeline

A production-ready transcription and diarization pipeline with parallel processing.

Features

• Parallel Processing: Runs Whisper transcription and NeMo diarization simultaneously
• Multiple Backends: Supports both faster-whisper and WhisperX
• Speaker Diarization: Uses NeMo MSDD models for accurate speaker identification
• Audio Source Separation: Optional vocal extraction using Demucs
• Punctuation Restoration: Automatic punctuation using deep learning models
• Memory Efficient: Proper GPU memory management and cleanup

Installation

pip install whisperx-nemo-pipeline

With constraints (recommended for production):

pip install whisperx-nemo-pipeline -c constraints.txt

Quick Start

from whisperx_nemo_pipeline import create_transcription_pipeline

# Create pipeline
pipeline = create_transcription_pipeline(
    audio_path="path/to/your/audio.wav",
    model_name="large-v2",
    device="cuda",  # or "cpu"
    stemming=True,  # Enable source separation
    backend="faster_whisper"  # or "whisperx"
)

# Process audio
transcript_path, srt_path, timing_info = pipeline.process()

print(f"Transcript saved to: {transcript_path}")
print(f"Subtitles saved to: {srt_path}")
print(f"Processing took: {timing_info['total_time']:.2f}s")

Advanced Usage

from whisperx_nemo_pipeline import TranscriptionPipeline, TranscriptionConfig

# Custom configuration
config = TranscriptionConfig(
    audio_path="path/to/audio.wav",
    model_name="large-v2",
    device="cuda",
    batch_size=8,
    language="en",  # or None for auto-detection
    stemming=True,
    suppress_numerals=False,
    backend="faster_whisper"
)

# Create pipeline with custom config
pipeline = TranscriptionPipeline(config)

# Process
transcript_path, srt_path, timing_info = pipeline.process()

Configuration Options

• audio_path: Path to input audio file
• model_name: Whisper model size ("tiny", "base", "small", "medium", "large-v2")
• device: Computing device ("cuda" or "cpu")
• batch_size: Batch size for inference (default: 4)
• language: Language code or None for auto-detection
• stemming: Enable audio source separation (default: True)
• suppress_numerals: Suppress numerical tokens (default: False)
• backend: "faster_whisper" or "whisperx"

Requirements

• Python 3.8+
• CUDA-capable GPU (recommended)
• See requirements.txt for full dependency list

License

MIT License