endman100/skill-router-nexus/asr-router

ASR Router

Treat this skill as the single interface for speech recognition. Keep downloading, editing, caption rendering, and business-specific QC in the calling skill.

Workflow

1. Read references/schema.md. Record the requested

profile, privacy policy, GPU availability, language, timing granularity, and whether fallback is allowed.

1. Select the first available method in the schema that satisfies every required

capability. A method that lacks required word timestamps, realtime support, privacy permission, credentials, or installed runtime is unavailable.

1. Before uploading media to an API, require privacy=allow_api. Stop after the

local methods when privacy=local_only.

1. Read only the selected method file and follow its invocation exactly:

• Qwen3-ASR — local GPU
• Whisper CLI — local
• Seed ASR — API
• Deepgram — API
• Paraformer — API
• OpenAI Whisper API — API
• Gladia — API
• Rev AI — API
• Wit.ai — API

1. Normalize the result to the schema. Preserve the raw provider artifact and

record the exact provider, model, command, source hash, and fallback trace.

1. On failure, record the error and try the next eligible method only when

fallback is allowed. Never silently cross from local processing to an API.

Invariants

• Use ASR output as recognized text unless a calling workflow supplies approved

display text. In that case, use ASR only as the timing substrate.

• Keep one recognition result per source hash, provider, model, and parameter

set. Do not mix timestamps or text from different runs.

• Preserve caller constraints such as preferred_provider=qwen3,

fallback_allowed=false, or production word-timestamp QC even when another method ranks higher globally.

• A calling skill may lock a provider only through Router input constraints. It

must not locate or execute a provider adapter itself.

• Report actual availability. A catalogue entry or uninstalled upstream bundle

is not an executable fallback.