yt-whisper-docker

MAIN REPOSITORY: https://git.digtvbg.com/ivangotoy/yt-whisper-docker

READ-ONLY MIRROR: https://github.com/ivangotoy/yt-whisper-docker.git

YouTube audio transcription with host yt-dlp, Docker, ffmpeg, and whisper.cpp.

The host wrapper downloads YouTube audio with host yt-dlp, including browser-cookie support, then runs Docker only for WAV conversion and transcription with the selected whisper.cpp backend.

Images

Default local image names:

• yt-whisper-docker:cpu - CPU backend
• yt-whisper-docker:amd - Vulkan backend for AMD / Mesa / /dev/dri
• yt-whisper-docker:nvidia - CUDA backend for NVIDIA

Optional registry builds are supported with:

REGISTRY=digtvbg.com:6000 ./scripts/build.sh amd

Requirements

• Linux or WSL2
• Docker with Buildx
• yt-dlp installed on the host
• For AMD / Vulkan: /dev/dri
• For NVIDIA / CUDA: NVIDIA driver and NVIDIA Container Toolkit

Build

Build CPU:

./scripts/build.sh cpu

Build AMD / Vulkan:

./scripts/build.sh amd

Build NVIDIA / CUDA:

./scripts/build.sh nvidia

Build all three serially:

./scripts/build.sh all

Useful build variables:

Variable	Default
REGISTRY	empty
IMAGE_NAME	yt-whisper-docker
UBUNTU_IMAGE	digtvbg.com:6000/home/ubuntu:latest
UBUNTU_VERSION	26.04
CUDA_VERSION	13.3.0
WHISPER_CPP_REF	master
YT_WHISPER_MODEL	large-v3
BUILD_JOBS	nproc
NVIDIA_BUILD_JOBS	24
CMAKE_CUDA_ARCHITECTURES	75;80;86;90

Run

Run CPU:

./scripts/run.sh cpu 'https://www.youtube.com/watch?v=VIDEO_ID'

Run AMD / Vulkan:

./scripts/run.sh amd 'https://www.youtube.com/watch?v=VIDEO_ID'

Run NVIDIA / CUDA:

./scripts/run.sh nvidia 'https://www.youtube.com/watch?v=VIDEO_ID'

Use a custom image tag:

YT_WHISPER_IMAGE=digtvbg.com:6000/yt-whisper-docker:amd ./scripts/run.sh amd URL

Cookies

The normal workflow uses host yt-dlp, so browser cookies stay on the host and are not baked into images.

Use a cookies file:

YT_WHISPER_COOKIES="$HOME/Downloads/www.youtube.com_cookies.txt" ./scripts/run.sh amd URL

Use browser cookies:

YT_WHISPER_COOKIES_FROM_BROWSER=firefox ./scripts/run.sh cpu URL
YT_WHISPER_COOKIES_FROM_BROWSER=chrome ./scripts/run.sh cpu URL
YT_WHISPER_COOKIES_FROM_BROWSER='chromium:/home/user/.config/chromium/Default' ./scripts/run.sh amd URL

If no cookie variable is set, the wrapper tries common local browser profiles automatically.

Runtime variables

Variable	Default	Purpose
YT_WHISPER_OUTPUT_DIR	./output	Host output directory
YT_WHISPER_LANG	auto-detected	Force language, for example bg, en, de
YT_WHISPER_THREADS	nproc inside the container	Inference threads passed to whisper-cli -t
YT_WHISPER_MAX_CONTEXT	0	Value passed to whisper-cli -mc
YT_WHISPER_KEEP_TEMP	0	Keep downloaded audio and WAV when set to 1
YT_WHISPER_YTDLP_FORMAT	bestaudio/best	Host yt-dlp format selector
YT_WHISPER_ALLOW_LANG_MISMATCH	0	Allow forced language mismatch when set to 1
YT_WHISPER_REQUIRE_GPU	1 for NVIDIA check	Require CUDA log verification for NVIDIA wrapper runs
YT_WHISPER_MODEL_FILE	/models/large-v3/ggml-large-v3.bin	Override model path inside the container

Output

Default output directory:

./output

Generated paths:

audio/
wav/
transcripts/
logs/

By default, downloaded audio and temporary WAV are deleted after successful transcription. Use YT_WHISPER_KEEP_TEMP=1 to keep them.

Verification status

CPU was built and runtime verified:

use gpu = 0
n_threads = 32 / 32

AMD / Vulkan was built and runtime verified on AMD Radeon RX 9070 XT:

Vulkan0 backend
n_threads = 32 / 32
output files owned by the host user

NVIDIA / CUDA was built and static-image verified:

yt-dlp present
ffmpeg present
whisper-cli present
model present
YT_WHISPER_THREADS support present
WHISPER_BACKEND=cuda present

NVIDIA runtime inference requires a host with NVIDIA GPU and NVIDIA Container Toolkit.