Whisper on AMD Windows

Run OpenAI's Whisper speech-to-text model on AMD GPUs on Windows with zero CUDA dependency. This guide covers two GPU acceleration paths plus CPU fallback, tested on RX 5700 XT (gfx1010 / RDNA1).

What is Whisper?

Whisper is OpenAI's open-source speech recognition model trained on 680,000 hours of multilingual audio. It transcribes audio to text with high accuracy across 99 languages and handles diverse audio conditions (accents, background noise, technical language).

Why AMD GPU?

No CUDA lock-in: RDNA cards work on Windows via DirectML (any DX12 GPU) or ROCm (RDNA1+ only)
Cost-effective: RX 6700 XT ~$300; equivalent NVIDIA RTX 3070 Ti $400–500
Sufficient VRAM: RX 5700 XT (8GB) → runs large model; RX 6600 XT (16GB) → runs large-v3 with headroom
Real transcription speedup: ~8–12x realtime on medium model vs ~0.5x CPU realtime

What This Repo Covers

Method	GPU Support	Windows	Difficulty	Speed	Notes
A: faster-whisper + DirectML	Any AMD DX12 (RDNA, Polaris, Vega)	✓	Easiest	8–12x realtime	Recommended — works on all AMD cards
B: whisper.cpp GPU	RDNA1+ only (ROCm via comfyui-rocm)	✓	Moderate	10–15x realtime	Requires ROCm environment from comfyui-amd-windows-setup
C: CPU fallback	Any	✓	Trivial	0.3–0.5x realtime	For testing or low-VRAM fallback

GPU Compatibility

AMD GPU Family	Architecture	gfx Code	Method A (DirectML)	Method B (ROCm)	Tested
RDNA4 (2025+)	RDNA4	gfx1201+	✓ Full	✓ Full	Not yet
RX 7000	RDNA3	gfx1100–gfx1102	✓ Full	✓ Full	Not yet
RX 6000	RDNA2	gfx1030–gfx1102	✓ Full	✓ Full	Community reports ✓
RX 5000	RDNA1	gfx1010–gfx1012	✓ Full	✓ Full	RX 5700 XT (gfx1010) ✓
RX Vega	Vega	gfx900/906	✓ DirectML only	✗ No ROCm	Not tested
RX Polaris	Polaris	gfx803	✓ DirectML only	✗ No ROCm	Not tested

DirectML works on any AMD GPU with DX12 support (Windows 10+). ROCm on Windows is limited to RDNA1+.

Quick Start

Prerequisites

Windows 10 or 11
Python 3.10, 3.11, or 3.12 (64-bit)
AMD GPU with ≥4GB VRAM (8GB+ recommended)
~500MB disk space (model cache)

Method A: faster-whisper + DirectML (Recommended)

Fastest setup, works on all AMD GPUs.

# Clone this repo
git clone https://github.com/ChharithOeun/whisper-amd-windows.git
cd whisper-amd-windows

# Create virtual environment
python -m venv venv
venv\Scripts\activate

# Install dependencies
pip install --upgrade pip
pip install faster-whisper onnxruntime-directml

# Verify GPU detection
python scripts/verify_gpu.py

# Transcribe audio
python scripts/transcribe.py path/to/audio.mp3

# Benchmark your GPU
python scripts/benchmark.py

Method B: whisper.cpp + ROCm (Advanced, RDNA1+ only)

For maximum performance on RDNA1+ cards using ROCm:

# Requires comfyui-amd-windows-setup environment
# See: https://github.com/ChharithOeun/comfyui-amd-windows-setup

git clone https://github.com/ChharithOeun/whisper-amd-windows.git
cd whisper-amd-windows

# Use the pre-configured ROCm environment
source /path/to/comfyui-rocm-env/bin/activate

pip install -r requirements-rocm.txt
python scripts/transcribe_rocm.py path/to/audio.mp3

Method C: CPU Fallback

If GPU acceleration isn't available:

pip install faster-whisper
python scripts/transcribe.py path/to/audio.mp3 --device cpu

Installation & Setup — Method A (Step by Step)

1. Install Python

Download Python 3.11 (64-bit Windows installer).

# Verify installation
python --version  # Should be 3.10, 3.11, or 3.12

2. Clone Repository

git clone https://github.com/ChharithOeun/whisper-amd-windows.git
cd whisper-amd-windows

3. Create Virtual Environment

python -m venv venv
venv\Scripts\activate

You should see (venv) in your prompt.

4. Upgrade pip

python -m pip install --upgrade pip

5. Install Core Dependencies

pip install faster-whisper onnxruntime-directml

faster-whisper: Optimized Whisper inference engine
onnxruntime-directml: ONNX Runtime with DirectML GPU support

6. Verify GPU Detection

python scripts/verify_gpu.py

Expected output on AMD GPU:

GPU Detection Report
====================
DirectML available: True
AMD GPU detected: True
Device: DirectML device 0 (AMD Radeon RX 5700 XT)
CUDA not found (expected on AMD)

7. Transcribe Audio

python scripts/transcribe.py examples/sample.mp3

Outputs:

sample.txt — Plain text transcript
sample.srt — Subtitle file (timestamp segments)

8. Benchmark Your Setup

python scripts/benchmark.py

Shows model size, device, speed (x realtime), VRAM usage.

Proof of Concept — Tested on RX 5700 XT

Transcription Speed

Benchmark on RX 5700 XT (8GB VRAM) with various Whisper models:

Model	Size	Speed	VRAM	Time for 60s Audio
tiny	39MB	~45x realtime	0.4GB	1.3s
base	140MB	~18x realtime	0.8GB	3.3s
small	466MB	~12x realtime	1.2GB	5.0s
medium	1.5GB	~8x realtime	2.4GB	7.5s
large	2.9GB	~4x realtime	4.2GB	15s

Baseline (CPU only): ~0.3x realtime on medium = 200s for 60s audio. Speedup: ~26x faster with RX 5700 XT.

Sample Output

Transcribing a 2-minute podcast episode (128 kbps MP3):

Input: podcast_episode_001.mp3 (2:34)
Model: medium
Device: DirectML
Total time: 19.2 seconds
Speed: ~8.0x realtime

Output: podcast_episode_001.txt
[Transcript of full 2:34 audio captured]

Output: podcast_episode_001.srt
00:00:00,000 --> 00:00:05,500
Welcome to the AI podcast, where we discuss...

00:00:05,500 --> 00:00:12,000
the latest developments in machine learning and open source.

Model Recommendations by GPU VRAM

Choose the largest Whisper model your GPU can fit:

GPU VRAM	Recommended Model	Use Case
2–4GB	`small` (466MB)	Real-time speech, live captions
4–6GB	`medium` (1.5GB)	Podcast, meeting transcription
6–8GB	`medium`–`large` (1.5–2.9GB)	High accuracy, document transcription
8GB+	`large`, `large-v3` (2.9GB)	Maximum accuracy across languages

Rule of thumb: Model should be ≤ 40% of your GPU VRAM to leave room for activations.

Troubleshooting

"No GPU detected" or "CUDA not found"

Cause: DirectML not detected or older drivers.

Fix:

# Update AMD GPU drivers
# Download from https://www.amd.com/en/support

# Verify DirectML is available
python -c "import onnxruntime; print(onnxruntime.get_available_providers())"
# Should show 'DmlExecutionProvider' in list

# If not, reinstall onnxruntime-directml:
pip uninstall onnxruntime-directml onnxruntime
pip install onnxruntime-directml

Out of Memory (OOM)

Cause: Model too large for GPU VRAM.

Fix:

Use smaller model: medium → small

Enable CPU offloading (slows down inference slightly):

python scripts/transcribe.py audio.mp3 --compute-type int8

Reduce batch size in transcribe.py

DirectML fails on CPU fallback

Cause: DirectML provider not registered.

Fix:

# Use CPU explicitly
python scripts/transcribe.py audio.mp3 --device cpu

Audio format not supported

Supported formats: MP3, WAV, M4A, FLAC, OGG, WMA

Convert with FFmpeg:

# Install FFmpeg: https://ffmpeg.org/download.html
ffmpeg -i unsupported.format -acodec libmp3lame -ab 192k output.mp3

ROCm (Method B) not detecting GPU

Cause: ROCm environment not sourced.

Fix:

# Must activate comfyui-rocm environment first
source /path/to/comfyui-rocm-env/bin/activate  # Linux WSL
# or on native ROCm on Windows (if HIP SDK installed):
set HIP_VISIBLE_DEVICES=0
rocminfo  # Verify GPU detection

Related Repositories

comfyui-amd-windows-setup — Complete guide to setting up ComfyUI with AMD GPU on Windows (ROCm + DirectML)
AMD-GPU-Windows-Setup — Baseline AMD GPU driver and ROCm setup for Windows

Keywords

whisper, speech-to-text, amd-gpu, directml, rocm, windows, gpu-acceleration, transcription, onnxruntime, faster-whisper, rdna, rdna2, rdna3, rx-5700-xt, rx-6700-xt, gpu-computing

Contributing

Found an issue? Have a faster-whisper optimization? Tested on another AMD GPU? Open an issue or PR:

Fork the repository
Create a branch: git checkout -b feature/gpu-xyz
Add your changes + test output (benchmark results, transcription samples)
Push and open a pull request

Please include:

Your GPU model and driver version
Python version and OS build
Benchmark output from scripts/benchmark.py
Any new methods or optimizations tested

License

MIT License — see LICENSE file.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files to use, modify, and distribute freely.

Support & Community

Issues & Questions: GitHub Issues
Discussions: GitHub Discussions
Buy Me a Coffee: Support development

Last tested: April 2026 | Python: 3.10–3.12 | Windows: 10/11 (build 19041+)

Name		Name	Last commit message	Last commit date
.github		.github
scripts		scripts
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
CONTRIBUTING.md		CONTRIBUTING.md
INSTALL.md		INSTALL.md
LICENSE		LICENSE
MANIFEST.md		MANIFEST.md
PROJECT_STRUCTURE.md		PROJECT_STRUCTURE.md
README.md		README.md
config.example.json		config.example.json
requirements.txt		requirements.txt
run.bat		run.bat
setup.py		setup.py

Latest commit

History

Folders and files

Repository files navigation

Whisper on AMD Windows

What is Whisper?

Why AMD GPU?

What This Repo Covers

GPU Compatibility

Quick Start

Prerequisites

Method A: faster-whisper + DirectML (Recommended)

Method B: whisper.cpp + ROCm (Advanced, RDNA1+ only)

Method C: CPU Fallback

Installation & Setup — Method A (Step by Step)

1. Install Python

2. Clone Repository

3. Create Virtual Environment

4. Upgrade pip

5. Install Core Dependencies

6. Verify GPU Detection

7. Transcribe Audio

8. Benchmark Your Setup

Proof of Concept — Tested on RX 5700 XT

Transcription Speed

Sample Output

Model Recommendations by GPU VRAM

Troubleshooting

"No GPU detected" or "CUDA not found"

Out of Memory (OOM)

DirectML fails on CPU fallback

Audio format not supported

ROCm (Method B) not detecting GPU

Related Repositories

Keywords

Contributing

License

Support & Community

About

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages