Claude-Transcription

April 19, 2026 · View on GitHub

End-to-end audio-to-transcript pipeline as a Claude Code plugin. Preprocess long-form recordings, transcribe via cloud or local engines, post-process into clean/structured/blog/summary formats, combine versions, and export.

Pipeline

audio ──► preprocess ──► transcribe ──► post-process ──► combine ──► export

Skills

Preprocessing

  • denoise — remove background noise. Cloud: Auphonic (default, cheapest), ElevenLabs Voice Isolator, Dolby.io. Local: DeepFilterNet (ML) or ffmpeg afftdn (non-ML).
  • truncate-silence — VAD-based silence removal (silero-vad or ffmpeg silenceremove).
  • normalize-format — stereo→mono, 16kHz resample, compress WAV to opus/mp3.
  • extract-speaker-samples — cluster unique voices in the audio, emit short per-speaker samples for user to label (feeds diarization).

Transcription

  • transcribe-gemini-raw — verbatim via gemini-transcription MCP.
  • transcribe-gemini-cleaned — filler words removed at transcription time.
  • transcribe-assemblyai — timestamped + diarization via AssemblyAI.
  • transcribe-whisper-local — offline transcription via local Whisper.
  • setup-whisper — one-time local Whisper installation.

Post-processing

  • clean-transcript — strip fillers (ums, likes, repetitions) from an existing transcript.
  • structure-transcript — add headers, logical sections, paragraph breaks.
  • transcript-to-blog — rewrite as a publishable blog post.
  • transcript-to-summary — executive summary + bullet highlights.
  • transcript-to-notes — structured meeting/study notes.

Compilation & Export

  • combine-versions — concatenate raw + cleaned + structured into one doc. Markdown (variant headers) or Typst → PDF (page numbers, variant in footer).
  • export-transcript — email, upload to Google Drive, or copy to clipboard.

Config

  • configure — onboarding: pick denoise + transcription providers, set default output directory, toggle cloud/local preference.

Defaults

  • Cloud over local for every skill that offers both. Override via configure or per-invocation.
  • Denoise: Auphonic (cheapest cloud option at ~$0.15/hr).
  • Transcription: Gemini (raw or cleaned).

See CLAUDE.md for config file schema and conventions.

Installation

claude plugins marketplace add danielrosehill/Claude-Code-Plugins
claude plugins install claude-transcription@danielrosehill

Then restart Claude Code.

External requirements

  • ffmpeg — preprocessing
  • python3 + uv — for skills that shell out to Python (DeepFilterNet, silero-vad, pyannote, whisper)
  • MCP servers: gemini-transcription, gws-personal (for Drive export)
  • API keys (set as env vars, names configurable): AUPHONIC_API_KEY, ELEVENLABS_API_KEY, ASSEMBLYAI_API_KEY, DOLBY_API_KEY

License

MIT