nemotron-asr-finetune

Nemotron Speech ASR Customization — Orchestration Skill

Note: "Nemotron Speech" is the public-facing name for what NVIDIA documents today as Riva / Riva NIM; the acoustic models are trained and fine-tuned with NVIDIA NeMo. Commands, config paths, imports, and doc URLs still use "Riva" / "NeMo" — the rename is brand-only. Do not rename them.

What This Skill Is

This is a high-level orchestration skill, not a step-by-step training manual. Its job, given a goal such as "I want to fine-tune ASR for my domain/language", is to:

1. Scope the problem (how much real audio, target eval set, latency/hardware budget, language/domain).
2. Choose the cheapest sufficient path — word boosting, n-gram LM fusion, or fine-tuning — and escalate only when quality falls short.
3. Delegate each stage to the right sub-skill (data generation, training, evaluation, deployment/optimization).
4. Answer cost/time/data questions along the way (how many hours to hit X% WER, synthetic vs real, L40S vs H100, expected cost).

It owns the plan and the routing; the sub-skills own the execution. When a needed sub-skill does not exist yet, this skill names it as a placeholder and gives interim guidance.

When to Use

Use for any request to make a Nemotron Speech / Riva ASR model work better on a specific domain or language — improving accuracy, reducing WER, adding a language, or planning a fine-tune. Start here even when the user names a specific technique: treat it as a candidate until the orchestration step confirms it is the cheapest sufficient path, then sequence the right sub-skills.

Orchestration Workflow

Run the loop below; each stage names the sub-skill it invokes. Full detail in references/workflow.md.

Stages 4–8 are the fine-tune path (data → NeMo train → NeMo eval → Riva deploy). Cheaper rungs (boosting, custom vocab, n-gram LM) take a shorter branch owned by a single sub-skill — don't force them through the full loop. See the branch-by-rung table in references/workflow.md (§3b).

Before training (Stage 5), run the pre-flight dataset quality check (§4a): verify/convert audio to 16 kHz, and, on user request, audit transcript quality by running a reference pretrained model and checking WER against the provided ground truth. See references/workflow.md §4a.

Also before training (Stage 5), run the pre-flight environment/dependency check (§4b): if no NeMo is provided, pull the latest main (local execution) or the latest published container tag (container execution). If a NeMo checkout/install is already provided, check its version, but still recommend switching to latest main (staleness risk) and ask the user — if they insist on the provided one, proceed with it and only revisit once a concrete version issue (e.g. unsupported functionality) is actually hit. See references/workflow.md §4b.

Throughout, answer the "along the way" questions (data volume, synthetic vs real, hours to reach a WER target, cost, GPU choice) — see references/planning-answers.md.

Sub-Skills This Skill Calls

Detailed registry, invocation, and handoff contracts in references/sub-skills.md.

If a sub-skill is unavailable, say so, give the interim guidance from the reference, and continue the plan.

Choosing The Path (cheapest first)

The scoping in Stage 3 selects the lowest-cost rung that can meet the target. Summary; full docs-grounded ladder in references/path-selection.md.

• Word boosting — a bounded set of known words/names/jargon. Two realizations that are different artifacts: pilot (NeMo) to prove lift offline via GPU-PB context biasing (nemo-speech-asr-finetune), or deploy (Riva) to ship it via runtime boosted_lm_score (nemotron-speech) — different score systems, don't reuse one for the other. Runtime, no training either way. See references/path-selection.md.
• Custom vocabulary / pronunciation — OOV or consistently mispronounced terms. Deploy-time. → Deployment sub-skill.
• N-gram (KenLM) LM — domain phrasing/word-sequences when you have text but little audio. Two realizations, built the same way but deployed differently by architecture: pilot (NeMo, NGPU-LM) to prove lift offline (nemo-speech-asr-finetune), or deploy (Riva) to ship it (nemotron-speech). For CTC, rebuild the pilot corpus into a Riva word-level LM — don't ship the pilot artifact as-is. For RNN-T/TDT, the opposite: hand the pilot's .nemo artifact to Riva unchanged, no rebuild. See references/path-selection.md.
• Fine-tune — real acoustic gaps (accents, noise, channel) with enough transcribed audio (NIM guide: 100+ h; ~10 h floor only if mixed to avoid catastrophic forgetting). Below ~10 h, do not recommend fine-tuning — recommend word boosting instead (severe catastrophic-forgetting/overfitting risk). → Research/Training.
• Train from scratch / cross-language transfer — a new language with no suitable checkpoint (last resort). → Research/Training.

Ordering and per-model support follow the NVIDIA Speech NIM ASR customization guide: https://docs.nvidia.com/nim/speech/latest/asr/customization/customization.html.

Key Principles

• Scope before you pick. Don't recommend fine-tuning before the discovery questions and a measured baseline.
• Cheapest sufficient path. Escalate rungs only when the current one provably can't hit the target; you may experiment on a cheap rung while presenting the full fine-tuning plan.
• Measure with a contract. Report normalized WER on the domain set plus an A/B forgetting check on a general set — never in-training logs alone.
• Verify before you train. Run the pre-flight dataset quality check before Stage 5: verify/convert audio to 16 kHz, and, when the user asks, audit ground-truth transcripts by comparing them to a reference model's WER. Don't train on an unresampled or unaudited-on-request dataset. See references/workflow.md §4a.
• Know your environment before you train. If no NeMo is provided, pull the latest main (or latest published container tag). If one is provided, check its version but still recommend latest main and ask the user first — a provided checkout always carries staleness risk. If they insist on keeping it, don't pull preemptively; only revisit once a concrete version issue (e.g. unsupported functionality) is actually hit. Record what was resolved in the ledger. See references/workflow.md §4b.
• Delegate, don't reimplement. Route execution to the sub-skills; keep this skill focused on the plan, sequencing, and cost/time/data answers.
• Real target-domain audio is the usual bottleneck. Prefer real data; use synthetic to fill measured gaps, kept separately weighted so it can be ablated.
• Consult the user before extra tuning cycles, and when a needed sub-skill is a placeholder.

Source of Truth

Limitations

• Orchestration only — execution happens in the sub-skills. Where a sub-skill is a placeholder, guidance is interim until it exists.
• GPU required for the training rungs; deployment/serving is owned by the nemotron-speech sub-skill.
• Model names, config paths, flags, and per-model feature support drift across NeMo/Riva releases — verify against the support matrix and the current checkout.
• Public branding is "Nemotron Speech"; commands, imports, config paths, and doc URLs still use "Riva" / "NeMo" — do not rename.