Faster-Whisper is a reimplementation of OpenAI's Whisper model using CTranslate2, a fast inference engine for Transformer models. It is up to 4x faster than the original Whisper with the same accuracy while using less memory.

Supports NVIDIA GPU (CUDA 12 + cuDNN 9), CPU with INT8 quantization

Requirements

• Python 3.9 or higher
• Linux (Ubuntu 20.04+, Debian 11+, Fedora, Arch, etc.)
• NVIDIA GPU with CUDA 12 drivers (optional)

Step-by-Step Installation on Linux

1️⃣ Check Python

Most Linux distributions come with Python 3. Verify:

python3 --version
pip3 --version

If not installed:

# Ubuntu/Debian
sudo apt update && sudo apt install -y python3 python3-pip python3-venv

# Fedora
sudo dnf install python3 python3-pip

# Arch Linux
sudo pacman -S python python-pip

2️⃣ (Recommended) Create a Virtual Environment

python3 -m venv whisper-env
source whisper-env/bin/activate

3️⃣ Install Faster-Whisper

pip install faster-whisper

4️⃣ GPU Acceleration — NVIDIA CUDA (Optional)

Option A — Install via pip (recommended)

pip install nvidia-cublas-cu12 nvidia-cudnn-cu12==9.*

export LD_LIBRARY_PATH=$(python3 -c "import os, nvidia.cublas.lib, nvidia.cudnn.lib; print(os.path.dirname(nvidia.cublas.lib.__file__) + ':' + os.path.dirname(nvidia.cudnn.lib.__file__))")

Add to your ~/.bashrc:

echo 'export LD_LIBRARY_PATH=$(python3 -c "import os, nvidia.cublas.lib, nvidia.cudnn.lib; print(os.path.dirname(nvidia.cublas.lib.__file__) + ":" + os.path.dirname(nvidia.cudnn.lib.__file__))")' >> ~/.bashrc

Option B — Docker

docker run --gpus all -it --rm nvidia/cuda:12.3.2-cudnn9-runtime-ubuntu22.04
pip install faster-whisper

5️⃣ Test the Installation

Create test_whisper.py:

from faster_whisper import WhisperModel

# CPU mode
model = WhisperModel("tiny", device="cpu", compute_type="int8")

# GPU mode (if you have NVIDIA)
# model = WhisperModel("tiny", device="cuda", compute_type="float16")

segments, info = model.transcribe("audio.mp3", beam_size=5)

print(f"Detected language: {info.language} (probability: {info.language_probability})")
for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

Run it:

python3 test_whisper.py

🧠 Available Models

Linux Tips

• No GPU? Use device="cpu" with compute_type="int8" — excellent performance
• With NVIDIA GPU? Use device="cuda" with compute_type="float16" — up to 4x faster
• Batch transcription: Use BatchedInferencePipeline(model=model).transcribe("audio.mp3", batch_size=16) for long audio files
• VAD filter: Enable with vad_filter=True to skip silence automatically
• Word-level timestamps: Add word_timestamps=True for per-word timing

Batch transcription example:

from faster_whisper import WhisperModel, BatchedInferencePipeline

model = WhisperModel("large-v3", device="cuda", compute_type="float16")
batched_model = BatchedInferencePipeline(model=model)
segments, info = batched_model.transcribe("podcast.mp3", batch_size=16)

for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

References

• Official repo: https://github.com/SYSTRAN/faster-whisper
• CTranslate2 docs: https://github.com/OpenNMT/CTranslate2/
• Whisper models: https://github.com/openai/whisper

Posted by SupportDev — Questions? Ask right here!