Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
109 lines
4.2 KiB
Markdown
109 lines
4.2 KiB
Markdown
# Built-in recipes
|
|
|
|
Recipes are ready-to-run compositions of existing audio.cpp models. Unlike
|
|
`--pipeline`, a recipe does not require a JSON workflow file. The recipe defines
|
|
the stages and their data flow; the caller only selects one model for each role
|
|
and supplies the normal task inputs.
|
|
|
|
List the available recipes; the server response also includes their model roles:
|
|
|
|
```bash
|
|
audiocpp_cli --list-recipes
|
|
curl http://127.0.0.1:8080/v1/recipes
|
|
```
|
|
|
|
## Any TTS voice cloning
|
|
|
|
`any-tts` generates speech with an offline TTS model and converts each generated
|
|
chunk to the supplied reference voice with Tone Color VC. The two sessions are
|
|
created once and reused for every text chunk.
|
|
|
|
```bash
|
|
audiocpp_cli --recipe any-tts --backend cuda --log \
|
|
--recipe-model tts=/path/to/kokoro.gguf \
|
|
--recipe-family tts=kokoro_tts \
|
|
--recipe-model converter=/path/to/tone-color-vc.gguf \
|
|
--recipe-family converter=tone_color_vc \
|
|
--voice-id am_michael \
|
|
--voice-ref reference.wav \
|
|
--text "The next train leaves in ten minutes." \
|
|
--request-option converter.seed=42 \
|
|
--out cloned.wav
|
|
```
|
|
|
|
`text_chunk_size` and `text_chunk_mode` control recipe-level text chunking.
|
|
Unprefixed request options and `tts.*` options go to the TTS stage;
|
|
`converter.*` options go to Tone Color VC. Role-prefixed options override shared
|
|
options with the same name.
|
|
Prefix CLI load and session option keys with their role too, retaining any
|
|
model-family prefix after the role prefix.
|
|
|
|
## Cross-language dubbing
|
|
|
|
`cross-language-dubbing` translates input speech and sends the translated text
|
|
directly to a target-language TTS model. Canary 180M Flash is a compact choice
|
|
for its supported English, German, Spanish, and French pairs. OWSM and OWSM-CTC
|
|
expose the same translation interface for broader language coverage.
|
|
|
|
```bash
|
|
audiocpp_cli --recipe cross-language-dubbing --backend cuda --log \
|
|
--recipe-model translator=/path/to/canary-180m-flash-f32.gguf \
|
|
--recipe-family translator=canary_asr \
|
|
--recipe-model tts=/path/to/kokoro.gguf \
|
|
--recipe-family tts=kokoro_tts \
|
|
--audio german-speech.wav \
|
|
--request-option translator.language=de \
|
|
--request-option translator.target_language=en \
|
|
--voice-id am_michael \
|
|
--out english-dub.wav \
|
|
--text-out translation.txt
|
|
```
|
|
|
|
Prefix translation options with `translator.*`. Unprefixed request options and
|
|
`tts.*` options go to the TTS stage. The translation target is required because
|
|
omitting it would turn the recipe back into same-language transcription and
|
|
re-speaking. Translator and TTS language identifiers are model-specific, so
|
|
the recipe does not infer one from the other.
|
|
|
|
## Server
|
|
|
|
Recipes use the default legacy server runtime, not the experimental parallel
|
|
runtime. They use model IDs from the normal server configuration, so model
|
|
loading, lazy loading, locking, and unloading retain their existing behavior.
|
|
The request maps recipe roles to those IDs:
|
|
|
|
```bash
|
|
curl http://127.0.0.1:8080/v1/recipes/run \
|
|
-H "Content-Type: application/json" \
|
|
-d '{
|
|
"recipe": "cross-language-dubbing",
|
|
"models": {"translator": "canary", "tts": "kokoro"},
|
|
"request": {
|
|
"audio": "input.wav",
|
|
"voice_id": "am_michael",
|
|
"options": {
|
|
"translator.language": "de",
|
|
"translator.target_language": "en"
|
|
}
|
|
}
|
|
}'
|
|
```
|
|
|
|
Set `"stream": true` to receive recipe output events and the final result over
|
|
SSE. Recipe components must be configured for the offline task and mode shown by
|
|
`GET /v1/recipes`.
|
|
|
|
## Adding a recipe plugin
|
|
|
|
Recipe implementations are isolated source plugins under
|
|
`app/recipe/plugins/`. A plugin implements `IRecipe` and declares one
|
|
`RecipePluginRegistration` factory in its own translation unit. The build
|
|
discovers plugin sources in that directory, so adding one does not require
|
|
editing the recipe registry, CLI, server, or another recipe.
|
|
|
|
The `IRecipeModelExecutor` boundary gives every plugin the same model execution
|
|
surface. `roles()` declares the required task and mode for each component, and
|
|
`run()` connects their in-memory `TaskRequest` and `TaskResult` values. Use
|
|
`options_for_recipe_role()` to isolate role-prefixed options. Source plugins are
|
|
compiled into the CLI and server; loading arbitrary binary plugins at runtime
|
|
is intentionally outside this interface.
|