nemo-speech-asr-finetune

name	nemo-speech-asr-finetune
description	Guide NeMo Speech users through ASR fine-tuning with container setup and Lhotse training.

NeMo Speech ASR Fine-Tuning

Use this skill when a user wants to fine-tune a NeMo Speech ASR model, choose a checkpoint, adapt a tokenizer, configure Lhotse dataloading, train, average checkpoints, or evaluate a fine-tuned ASR .nemo checkpoint. Also use it for post-run refinement planning after fine-tuning.

Default posture:

• Use the NeMo container unless the user explicitly asks for local execution.
• Prefer Lhotse for train and validation dataloaders.
• Use trainer.max_steps, not trainer.max_epochs.
• Use val_wer as the checkpoint monitor for validation.
• By default, evaluate WER without capitalization and punctuation effects. Change that only when the user explicitly asks for raw/cased/punctuated scoring.
• Report final quality from standalone evaluation, not only in-training validation logs.

Staged Workflow

Load only the reference file needed for the current stage:

1. Setup and checkpoint selection: read references/setup-checkpoints.md.
2. Data prep, transcript-style preflight, Lhotse, bucketing, validation dataloader, and blends: read references/data-lhotse.md.
3. Architecture detection, tokenizer changes, and AED/Canary multitask metrics: read references/architecture-tokenizer-metrics.md.
4. Training, checkpoint averaging, and evaluation: read references/training-evaluation.md and, when reporting WER, references/evaluation-style-contract.md.
5. Post-run refinement, error analysis, curriculum, and general-vs-domain evaluation: read references/refinement-iteration.md.

If the user explicitly asks for parallel/sub-agent work, split the work by these same stages. Keep each agent scoped to one stage and have the main agent integrate the final command/config.