WhisperX Replicate

A Cog-based deployment of WhisperX for German speech-to-text transcription using the faster-whisper-large-v3-turbo model.

Overview

This repository packages WhisperX as a Replicate-compatible model, enabling easy deployment and inference via Cog. It uses:

• WhisperX (i4ds fork) for transcription with VAD (Voice Activity Detection)
• faster-whisper-large-v3-turbo model for fast, accurate German transcription
• Cog for containerization and deployment to Replicate

Prerequisites

• Cog installed
• NVIDIA GPU with CUDA 12.1 support
• Docker

Setup

1. Download the Model

First, download the model to your Hugging Face cache. You can use the helper script:

python get_models.py

This will download the model (i4ds/daily-brook-134) to your local Hugging Face cache.

2. Copy Model to Repository

Copy the cached model to the models/ directory:

./copy_models.sh

This creates the following structure:

models/
└── faster-whisper-large-v3-turbo/
    ├── config.json
    ├── tokenizer.json
    ├── vocabulary.json
    └── ...

3. Build the Cog Image

cog build

Usage

Run a Prediction

cog predict -i audio_file=@your_audio.mp3

Input Parameters

Parameter	Description	Default
audio_file	Audio file to transcribe	(required)
language	Language (fixed to German)	de
batch_size	Parallelization for transcription	8
temperature	Sampling temperature	0
vad_onset	VAD onset threshold	0.500
vad_offset	VAD offset threshold	0.363
align_output	Enable word-level timestamps	False
debug	Print timing/memory info	True

Example

cog predict -i audio_file=@interview.mp3

Output

The prediction returns:

• segments: Transcription in SRT subtitle format
• detected_language: The detected language code (e.g., de)

File Structure

├── cog.yaml              # Cog configuration (CUDA, Python, dependencies)
├── predict.py            # Main prediction class for Cog
├── requirements.txt      # Python dependencies
├── copy_models.sh        # Script to copy model from HF cache
├── get_vad_model_url.py  # Helper to download model
└── models/               # Local model directory
    └── faster-whisper-large-v3-turbo/

Deployment to Replicate

cog login
cog push r8.im/your-username/whisperx-german

Notes

• The model is hardcoded to German (de) transcription
• Uses float16 compute type for GPU efficiency
• VAD is enabled by default for better handling of speech segments