FFmpeg 8.0 (GPU + Whisper) — Docker build

This repository builds a feature-rich FFmpeg 8.0 container with:

• Hardware transcoding

  • NVIDIA NVENC/NVDEC (--enable-ffnvcodec) + CUDA LLVM toolchain
  • Intel VAAPI and oneVPL (--enable-vaapi, --enable-libvpl)

• Whisper speech-to-text audio filter (--enable-whisper, via whisper.cpp)

• A broad set of codecs, demuxers, filters, and protocols (see below)

• Redistributable licensing: uses GnuTLS (not OpenSSL) so there’s no --enable-nonfree; the final binary can be distributed under the GPL

The final image is slim: only ffmpeg, ffprobe, and the exact shared libraries they need are included.

⚠️ Vulkan is not enabled in this build. If you want Vulkan/libplacebo later, you’ll need to add the relevant dev packages and --enable-vulkan (and typically --enable-libplacebo, --enable-libshaderc) to your configure flags.

---

What’s inside (enabled features)

Video encoders/decoders

• AV1: libaom (enc/dec), librav1e (enc), libsvtav1 (enc), libdav1d (dec)
• H.264 / H.265: libx264, libx265, NVENC (e.g., h264_nvenc, hevc_nvenc)
• VP8/VP9: libvpx
• AVS2: libdavs2 (dec), libxavs2 (enc)
• MPEG-4 ASP: libxvid
• Theora: libtheora
• OpenH264: libopenh264
• JPEG 2000: libopenjpeg
• WebP / JPEG XL: libwebp, libjxl

Audio encoders/decoders & DSP

• AAC (native), MP3 (libmp3lame), Opus (libopus)
• Vorbis (libvorbis), Speex (libspeex), AMR-NB/AMR-WB (opencore-amr)
• MP2 (twolame)
• SoX Resampler (libsoxr), Rubber Band (time-stretch), Game Music (libgme)

Subtitles, text & filters

• Whisper audio-to-text filter (--enable-whisper)
• Subtitles/text: libass, freetype, fontconfig, harfbuzz, fribidi
• Stabilization: libvidstab
• Frei0r effects
• ARIB B-24 captions (libaribb24)
• Chromaprint (acoustic fingerprinting)

Protocols / I/O / misc

• SRT (GnuTLS build), RIST, SSH, ZeroMQ
• SDL2, OpenAL
• Bluray/DVD nav: libbluray, libdvdnav, libdvdread
• GnuTLS for TLS/HTTPS, Zlib, LZMA

GPU stacks

• NVIDIA: NVENC/NVDEC headers (nv-codec-headers) + CUDA LLVM
• Intel: VAAPI + oneVPL (libvpl)
• OpenCL is enabled; Vulkan is not (by design in this build)

The image also includes AviSynth+ headers only (no AvsCore build), which is sufficient for --enable-avisynth.

---

Build

Linux/macOS

export DOCKER_BUILDKIT=1
docker buildx build \
  --build-arg FFMPEG_VER=8.0 \
  --build-arg AVS_TAG=v3.7.5 \
  --build-arg WHISPER_TAG=v1.8.0 \
  --build-arg CUDAARCHS=86 \
  -t drjp81/ffmpegwhisper:latest.

PowerShell (Windows)

$env:DOCKER_BUILDKIT="1"
docker buildx build `
  --build-arg FFMPEG_VER=8.0 `
  --build-arg AVS_TAG=v3.7.5 `
  --build-arg WHISPER_TAG=v1.8.0 `
  --build-arg CUDAARCHS=86 `
  -t drjp81/ffmpegwhisper:latest.

---

Using the Whisper wrapper (extract_subs.sh)

This repo includes a small wrapper that runs FFmpeg with the Whisper audio filter to extract SRT subtitles next to your media file.

Script

#!/bin/sh
# extract_subs.sh  —  wrapper for the FFmpeg Whisper filter
# Usage:
#   extract_subs.sh "<inputfile>" "<modelpath>"
# Example:
#   extract_subs.sh "/path/to/Movie (2024).mkv" "/models/ggml-large-v3-turbo.bin"

set -eu

inputfile="$1"
modelpath="$2"

if [ -z "$inputfile" ] || [ -z "$modelpath" ]; then
  echo "Usage: $0 <inputfile> <modelpath>" >&2
  exit 1
fi

inputdir=$(dirname "$inputfile")
filename=$(basename "$inputfile")
filename_noext="${filename%.*}"

cd "$inputdir"
outputfile="${filename_noext}.srt"

echo "Running Whisper → ${outputfile}"
ffmpeg -i "$inputfile" -vn \
  -af "whisper=model=${modelpath}:language=en:queue=5:destination=${outputfile}:format=srt" \
  -f null -

Add the wrapper into the image

In the runtime stage of your Dockerfile, add:

# Bundle the wrapper
COPY extract_subs.sh /usr/local/bin/extract_subs.sh
RUN chmod +x /usr/local/bin/extract_subs.sh

Run it (Linux)

docker run --rm --gpus all \
  -v /path/to/videos:/videos \
  -v /path/to/models:/models \
  drjp81/ffmpegwhisper:latest\
  extract_subs.sh "/videos/Movie (2024).mkv" "/models/ggml-large-v3-turbo.bin"

Run it (PowerShell on Windows)

docker run --rm --gpus all `
  -v "D:\Videos:/videos" `
  -v "D:\Models:/models" `
  drjp81/ffmpegwhisper:latest`
  extract_subs.sh "/videos/Movie (2024).mkv" "/models/ggml-large-v3-turbo.bin"

Notes

• The model path must be readable in the container (bind-mount your models directory).
• You can change language=en to language=auto to auto-detect.
• queue=5 tunes the filter’s worker queue (increase on bigger GPUs).

---

Examples

Transcode with NVENC (NVIDIA):

docker run --rm --gpus all -v "$PWD:/work" drjp81/ffmpegwhisper:latest\
  ffmpeg -hwaccel cuda -i /work/in.mp4 -c:v h264_nvenc -b:v 5M -c:a aac /work/out.mp4

Extract subtitles only (wrapper):

docker run --rm --gpus all \
  -v "$PWD:/work" -v "$HOME/models:/models" drjp81/ffmpegwhisper:latest\
  extract_subs.sh "/work/in.mp3" "/models/ggml-large-v3-turbo.bin"

---

Runtime expectations

• CUDA driver libs (libcuda.so.1, libcudart.so.12, libcublas.so.12, …) are not shipped in the runtime image; they come from the host (via --gpus all).
• ffmpeg / ffprobe are patched with an RPATH to load /opt/ffmpeg/lib, and the runtime sets /etc/ld.so.conf.d/ffmpeg.conf to the same path.

---

Troubleshooting

• ERROR: whisper >= 1.7.5 not found using pkg-config during build Ensure the builder stage installs nvidia-cuda-toolkit (and optionally nvidia-cuda-dev). The FFmpeg configure probe links against libggml-cuda.so and needs CUDA libs present at build time.

• libcuda.so.1 not found at runtime Run with --gpus all and have the NVIDIA drivers/toolkit installed on the host.

• VAAPI not working Pass through the device: --device /dev/dri:/dev/dri and ensure the user has permission to access /dev/dri/renderD128 on the host.

---

License

• FFmpeg is built with GPL options and GnuTLS (no OpenSSL; no nonfree).
• This repo contains no FFmpeg source; it fetches official releases and builds against distro-provided libraries.
• Non-commercial, attribution JpSoftworks Inc. (Quebec Canada)

---

Optional: verify build

Inside the container:

ffmpeg -hide_banner -buildconf | sed -n '1,120p'
ffmpeg -filters | grep -i whisper || true
ffmpeg -hwaccels