---

name: components-metrics description: "Guides SpeechBrain model components, decoders, losses, metric statistics, WER scoring, checkpoints, Pretrainer, streaming helpers, and component-level troubleshooting." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

SpeechBrain components and metrics

Use this sub-skill when the task is to choose, wire, debug, or test SpeechBrain neural components, decoders, metrics, checkpoints, pretrained-weight loading, streaming helpers, or scoring utilities.

Route map

Task	Read/run
Select neural modules, lobes, feature wrappers, decoders, or losses.	references/components-and-decoders.md
Save/recover checkpoints or load pretrained weights into modules.	references/checkpoints-and-pretraining.md
Compute WER/PER/CER, metric stats, classification/binary metrics, or scoring summaries.	references/metrics-and-scoring.md; run scripts/compute_wer_cli.py.
Work on streaming/chunked ASR or streaming feature wrappers.	references/streaming.md
Diagnose shape, checkpoint, metric, decoder, or streaming failures.	references/troubleshooting.md

Common component workflow

1. Keep model/module construction in HyperPyYAML when possible.
2. Use recipe Brain hooks to wire forward/objective/metric behavior.
3. Validate tensor shapes and relative lengths before decoder or metric calls.
4. Use Checkpointer for local experiment state and Pretrainer for fetching/loading pretrained model files.
5. Use metric stats classes inside on_stage_start / compute_objectives / on_stage_end patterns.

Useful bundled script

python scripts/compute_wer_cli.py ref.txt hyp.txt --mode strict

The script is adapted from SpeechBrain's WER utility pattern and uses the installed package only.