Automatic subtitle generation for your media using whisper.cpp.
Runs a webserver that upon receiving a webhook/post with a file path will use whisper.cpp to generate subtitles for your media and output them as srt files. Supports webhooks from Radarr & Sonarr.
Whisper.cpp is resource intensive, and go-subgen stores the stripped audio in memory instead of saving it to the filesystem. If you have large media files and/or use large models you will need lots of ram.
Note
I am in the process of substantially rewritting go-subgen to provide support for many of the below TODO features and improve code quality. This work is being done in the v1-dev and v1-dev-web branches, which are mostly functional.
- Sonarr & Radarr webhook support
- basic Tautulli webhook support
- Subtitle filename templating (see subfile-name-templating)
- File queueing
Create an issue or Discussion if you have other feature requests or comments.
- Further software integrations
- Persistent Queue
- Translation and more advanced media checking (don't run if file already has subs, for example)
- A basic web ui to queue transcription and view queued/in progress tasks.
- Filesystem watching support
- Integration into Bazarr as a provider (Not sure how to go about this, please open an issue if you have any ideas)
- CLBLAST & CUBLAS support for GPU acceleration #39
- Track ggerganov/whisper.cpp#1261
- Track ggerganov/whisper.cpp#1060 (Not present in Go binding yet)
- Distributed ASR?
Important
If you encounter illegal instruction errors when the model attempts to run, you may need to build the image locally. See Locally built image. The image currently requires your CPU support avx2, avx, and sse3.
version: "3.8"
services:
subgen:
image: ghcr.io/khakers/go-subgen:latest
restart: unless-stopped
ports:
- "8095:8095"
volumes:
# Path must be identical to the path in your other services
- /path/to/your/media:/media
# Models will be downloaded to this directory
- models:/models
environment:
- MODEL_TYPE=base_en
volumes:
models:In many cases you may want or even need to build the image locally. Building locally can give you better optimizations
for your specific hardware.
Unfortunately, this means you will need to manually trigger rebuilds (docker compose build) and change versions
yourself instead of letting a service like watchtower pull newer images for you.
version: "3.8"
services:
subgen:
build:
context: https://github.com/khakers/go-subgen.git#v0.1.0
restart: unless-stopped
ports:
- "8095:8095"
volumes:
# Path must be identical to the path in your other services
- /path/to/your/media:/media
# Models will be downloaded to this directory
- models:/models
environment:
- MODEL_TYPE=base_en
volumes:
models:GET
/healthcheck
Returns 200 if the server is running
POST
/webhooks/tautulli
Designed around being compatible with subgens Tautulli webhook and accepts the same json payload
However, go-subgen currently only uses 'file' and ignores the rest of the json data.
{
"event": "",
"file": "{file}",
"filename": "{filename}",
"mediatype": "{media_type}"
}POST
/webhooks/generic
A very basic post endpoint that accepts a json array of file paths
{
"files": [
"/path/to/file.mp4",
"/path/to/other/file.mkv"
]
}POST
/webhooks/radarr
Accepts Radarr formatted webhooks
Example webhook configuration
This example only sends the notification on series that have the 'whisper' tag, allowing you to only transcribe series that need it.
POST
/webhooks/sonarr
Accepts Sonarr formatted webhooks
Example webhook configuration
This example only sends the notification on series that have the 'whisper' tag, allowing you to only transcribe series that need it.
Go-Subgen allows you to configure how subtitle files are name using Go templates.
The default filename templates is as follows:
{{.FileName}}.subgen.{{.Lang}}.{{.FileType}}
You can set your own template by setting the environment variable SUBTITLE_NAME_TEMPLATE. The template is created
using the struct below. You can use any of the variables provided and any features of the Go templating system, but keep
in mind no escaping is applied to the result. Additionally, FileHash is not a SHA hash. It is a hash generated
using imohash
which hashes only portions of the file using murmur3
type SubtitleTemplateData struct {
FilePath string
FileName string
Lang string
FileHash string
FileType string
ModelType string
}| Environment Variable | Type | Default | Description |
|---|---|---|---|
| MODEL_TYPE | Model | base_en | Whisper.cpp Model version tiny_en, tiny base_en, base, small_en, small, medium_en, medium, large_v1, large_v2, large_v3 large |
| TARGET_LANG | string | en | |
| IGNORE_IF_EXISTING | bool | true | |
| MAX_CONCURRENCY | uint | 1 | |
| MODEL_DIR | string | /models/ | |
| LOG_LEVEL | log.Level | info | |
| VERIFY_MODEL_HASH | bool | true | Verify that the downloaded model mashes the expected hash |
| PORT | uint8 | 8095 | Web server port |
| SUBTITLE_NAME_TEMPLATE | string | "{{.FileName}}.subgen.{{.Lang}}.{{.FileType}}" | See Subfile-name-templating |
| WHISPER_CONF_THREADS | uint | Number of threads to run Whisper.cpp on | |
| WHISPER_CONF_WHISPER_SPEEDUP | bool | ||
| WHISPER_CONF_TOKEN_THRESHOLD | float32 | ||
| WHISPER_CONF_TOKEN_SUM_THRESHOLD | float32 | ||
| WHISPER_CONF_MAX_SEGMENT_LENGTH | uint | ||
| WHISPER_CONF_MAX_TOKENS_PER_SEGMENT | uint | ||
| FILE_PERMISSIONS_UID | uint | 0 | UID of the saved subtitle file |
| FILE_PERMISSIONS_GID | uint | 0 | GID of the saved subtitle file |
Inspired by McCloudS/subgen but written in Go and for more service agnostic (not designed around plex) usage.
![dependabot[bot] avatar](https://avatars.githubusercontent.com/in/29110?v=4 "dependabot[bot]")
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent