audio-tools-workflow

name

audio-tools-workflow

description

Use the RUDI audio-tools stack to transcribe, enrich, index, and query speech from local audio/video files, direct media URLs, supported video-page URLs, or base64 media input. Trigger when the user asks to process audio, video-linked speech, voice memos, meeting recordings, interviews, podcasts, clips, transcripts, transcript search, transcript enrichment, or audio-derived notes with the audio-tools stack.

Audio Tools Workflow

Use the local audio-tools MCP stack for speech-to-text workflows. The stack provides audio_transcribe, audio_enrich, audio_sync, audio_stats, and audio_query.

Workflow

1. Identify the source shape:
   • local file path: pass file
   • HTTP(S) direct media or supported video page: pass url
   • base64 media: pass data and filename
2. Call audio_transcribe first. It writes JSON and Markdown outputs under ~/.rudi/output/audio-tools/transcripts/YYYY/MM/DD/.
3. Preserve the returned transcript JSON path. Use it for enrichment and indexing.
4. Call audio_enrich only when the user needs metadata such as title, summary, tags, topics, people, sentiment, keywords, or action items.
5. Call audio_sync after transcription or enrichment when query/search/index results matter.
6. Use audio_stats for a quick collection overview.
7. Use audio_query for focused analysis across the SQLite transcript index.

Guardrails

• Treat transcripts as private user data.
• Do not include credentials, API keys, or raw secret values in prompts, outputs, config files, or stack packages.
• Only pass HTTP(S) URLs to URL input. The stack rejects other URL schemes.
• Prefer read-only SELECT queries with audio_query. Do not run destructive SQL unless the user explicitly asks for database modification.
• If enrichment fails because no prompt template is configured, report that transcription completed and name the missing configuration requirement.
• If a private source requires authorization, use the host connector to obtain the source file or URL before invoking the stack.
• If audio_transcribe reports a missing binary or Whisper model, treat that as a runtime provisioning issue in the MCP/plugin environment. Do not assume a successful manual shell command means the MCP server can see the same dependency.

Useful Files

• Connector hints: connectors/host-provided.json
• Environment and secret requirements: secrets/required.json
• MCP launch contracts: mcp/servers.json
• Transcript JSON schema: outputs/schemas/transcript-record.schema.json
• Query examples: references/query-examples.md