ragedunicorn/whisper repository overview

⁠Whisper Docker Image

Whisper speech recognition with whisper-ctranslate2⁠ (faster-whisper/CTranslate2) - transcribe and translate audio and video, on CPU or NVIDIA GPU.

⁠Quick Start

# Pull latest version (cpu variant)
docker pull ragedunicorn/whisper:latest

# Or the CUDA variant / a specific version
docker pull ragedunicorn/whisper:latest-cuda
docker pull ragedunicorn/whisper:0.5.7-cpu-1

# Transcribe a file (model is cached in a named volume)
docker run --rm -v $(pwd):/tmp/workdir -v whisper-models:/home/whisper/.cache/huggingface \
  ragedunicorn/whisper:latest video.mp4 --model medium --compute_type int8 --output_dir .

⁠Features

• 🚀 Fast: faster-whisper/CTranslate2 transcribes up to 4x faster than openai/whisper at the same accuracy
• ⚡ Two variants: portable CPU image (latest) and GPU-accelerated CUDA 12 image (latest-cuda)
• 🎥 Video files work directly: mp4/mkv audio is decoded by PyAV - no separate ffmpeg step
• 📦 Model cache: models download once into a mountable ~/.cache/huggingface volume
• 🔒 Non-root: Runs as the unprivileged whisper user by default
• 🏗️ Multi-platform: cpu variant supports linux/amd64 and linux/arm64 (cuda is amd64-only)

⁠Usage Examples

⁠Transcribe a video file

docker run --rm -v $(pwd):/tmp/workdir -v whisper-models:/home/whisper/.cache/huggingface \
  ragedunicorn/whisper:latest video.mp4 --model medium --compute_type int8 --output_dir .

⁠Generate SRT subtitles

docker run --rm -v $(pwd):/tmp/workdir -v whisper-models:/home/whisper/.cache/huggingface \
  ragedunicorn/whisper:latest video.mp4 --model small --output_format srt --output_dir .

⁠Translate any language to English

docker run --rm -v $(pwd):/tmp/workdir -v whisper-models:/home/whisper/.cache/huggingface \
  ragedunicorn/whisper:latest interview.wav --task translate --model small --output_dir .

⁠GPU transcription (NVIDIA, CUDA 12 driver required)

docker run --rm --gpus all -v $(pwd):/tmp/workdir -v whisper-models:/home/whisper/.cache/huggingface \
  ragedunicorn/whisper:latest-cuda video.mp4 --model large-v3 --device cuda --output_dir .

Not included: live microphone transcription (needs an audio device) and speaker diarization (needs PyTorch).

⁠Tags

This image uses semantic versioning with a variant infix:

Format: {whisper_ctranslate2_version}-{cpu|cuda}-{build_number}

⁠Version Examples

• latest - Most recent stable release (cpu variant)
• latest-cuda - Most recent stable release (cuda variant)
• 0.5.7-cpu-1 - whisper-ctranslate2 0.5.7, CPU variant, build 1
• 0.5.7-cuda-1 - whisper-ctranslate2 0.5.7, CUDA variant, build 1
• 0.5.7-cpu-2 - Rebuild of the same version (base image bump, fixes)

When updates are available through automated dependency management, new releases are created with appropriate version tags.

⁠Links

• GitHub: https://github.com/RagedUnicorn/docker-whisper⁠
• Issues: https://github.com/RagedUnicorn/docker-whisper/issues⁠
• Releases: https://github.com/RagedUnicorn/docker-whisper/releases⁠

⁠License

MIT License - See GitHub repository⁠ for details.