---

name: recipe-training description: "Guides SpeechBrain recipe and template training workflows, HyperPyYAML overrides, Brain loops, debug/test-only runs, recipe catalogs, and CPU/GPU/DDP launch choices." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

SpeechBrain recipe training

Use this sub-skill when the task is to run, adapt, debug, or create SpeechBrain recipes or templates. This includes train.py, hparams.yaml, HyperPyYAML, Brain, RunOptions, dataset preparation, debug flags, and distributed training.

Core route

1. Read references/recipe-workflows.md to understand recipe layout, data prep, output folders, and command construction.
2. Read references/hyperpyyaml-and-run-options.md when editing hparams, placeholders, overrides, runtime options, or Brain code.
3. Read references/native-examples.md to map tiny integration examples and recipe CSV debug rows to verification ideas.
4. Read references/distributed-and-debugging.md for CPU/GPU debug, DDP, and multi-node launches.
5. Read references/troubleshooting.md for slow downloads, missing extras, bad overrides, and device failures.
6. Use scripts/inspect_recipe_catalog.py to inspect a tests/recipes/*.csv catalog and build a debug command without running the recipe.

Common command shapes

Local recipe run:

cd recipes/<dataset>/<task>/<model>
python train.py hparams/train.yaml --data_folder /path/to/data --device cpu

Debug a cataloged recipe from repository root:

python sub-skills/recipe-training/scripts/inspect_recipe_catalog.py \
  --csv /path/to/recipe-catalog.csv --row 2 --print-command --device cpu

DDP single-node pattern:

cd recipes/<dataset>/<task>/<model>
torchrun --standalone --nproc_per_node=4 train.py hparams/train.yaml

Safety rules

• Dataset preparation scripts can download or rewrite large datasets. Use --skip_prep=True when using pre-existing fixtures or when data prep is out of scope.
• Recipe-specific extra_requirements.txt files can install large or conflicting packages. Install only the selected recipe's extras.
• HyperPyYAML can construct Python objects. Treat untrusted recipe YAML as code.
• CPU debug runs prove control/data flow only; they do not verify full recipe performance or CUDA throughput.