Audio Transcription Workflow

Use this skill when the user wants to convert one or more local audio files into cleaned markdown transcripts using the same core workflow as a legacy voice-note transcription bot path, but without Telegram, webhooks, or queues.

What this skill does

For each input audio file, the bundled script:

1. tries local Whisper first
2. falls back to OpenAI transcription (gpt-4o-mini-transcribe by default) if Whisper fails
3. runs OpenRouter cleanup with openai/gpt-5-nano unless cleanup is explicitly skipped or the API key is unavailable
4. writes a markdown file with a YAML metadata header and the cleaned transcript body
5. writes a transcription_manifest.json summary into the output directory

Prefer this skill over simpler transcription paths when

Use this skill instead of a basic one-shot transcription skill when the user wants any of the following:

• the same workflow as the old VoiceNote Bot
• markdown output files, not just inline text
• batch transcription of many files or a folder
• cleanup/editing after raw STT
• voice notes placed on disk rather than sent through Telegram

Inputs this skill accepts

• a single audio file
• multiple audio files
• a folder containing audio files
• a glob such as /path/to/*.m4a
• recursive folder traversal with --recursive

Supported formats:

• .oga, .ogg, .mp3, .m4a, .wav, .webm, .mp4, .mpeg, .mpga

Output format

Each transcript is saved as a .md file with a metadata header like this:

yaml
---
source_file: "note-01.oga"
source_path: "/absolute/path/note-01.oga"
generated_at: "2026-05-28T10:00:00+00:00"
workflow: "local_whisper_primary_openai_fallback_openrouter_cleanup"
language_requested: "auto"
transcription_provider: "whisper"
transcription_model: "local-whisper"
cleanup_applied: true
cleanup_provider: "openrouter"
cleanup_model: "openai/gpt-5-nano"
warnings: []
---

The markdown body below the header is the transcript the user should read or reuse.

Bundled script

Use this script:

bash
python3 ./skills/audio-transcription-workflow/scripts/transcribe_audio_to_markdown.py ...

Common commands

Single file

bash
python3 ./skills/audio-transcription-workflow/scripts/transcribe_audio_to_markdown.py \
  /path/to/voice-note.oga \
  --output-dir /path/to/output

Multiple files

bash
python3 ./skills/audio-transcription-workflow/scripts/transcribe_audio_to_markdown.py \
  /path/to/a.oga /path/to/b.m4a /path/to/c.wav \
  --output-dir /path/to/output

Folder

bash
python3 ./skills/audio-transcription-workflow/scripts/transcribe_audio_to_markdown.py \
  /path/to/audio-folder \
  --output-dir /path/to/output

Recursive folder

bash
python3 ./skills/audio-transcription-workflow/scripts/transcribe_audio_to_markdown.py \
  /path/to/audio-folder \
  --recursive \
  --output-dir /path/to/output

Glob input

bash
python3 ./skills/audio-transcription-workflow/scripts/transcribe_audio_to_markdown.py \
  '/path/to/audio/*.m4a' \
  --output-dir /path/to/output

Raw transcript only

If the user explicitly wants no cleanup:

bash
python3 ./skills/audio-transcription-workflow/scripts/transcribe_audio_to_markdown.py \
  /path/to/audio-folder \
  --output-dir /path/to/output \
  --skip-cleanup

Environment requirements

Expected environment variables:

• WHISPER_URL — optional, defaults to http://localhost:9000/asr
• OPENAI_API_KEY — optional but needed for transcription fallback
• OPENAI_TRANSCRIPTION_MODEL — optional, defaults to gpt-4o-mini-transcribe
• OPENROUTER_API_KEY — optional but needed for legacy-style cleanup

The bundled script tries to work for low-context agents too.

• Credentials such as OPENROUTER_API_KEY and OPENAI_API_KEY are resolved in this order: current process environment → a shell startup file such as ~/.bashrc via the shared credential loader → an optional legacy dotenv file pointed to by AUDIO_TRANSCRIPTION_LEGACY_ENV if you deliberately enable that fallback.
• WHISPER_URL stays simple: current process environment first, otherwise the script uses its built-in localhost default.

That means cleanup should usually work even when an agent has not been explicitly told where the OpenRouter key lives.

How to operate the skill

1. Resolve the user's input path(s).
2. Decide on an output directory.

• Prefer a user-specified output directory.
• Otherwise use a sensible local folder such as ./transcriptions near the current task context.

1. Run the bundled script.
2. Read the generated .md files if you need to summarise results back to the user.
3. Mention:

• where the markdown transcripts were saved
• whether any file used the OpenAI fallback
• whether cleanup was skipped or failed for any file
• where transcription_manifest.json was saved

Behaviour notes

• The script serialises Whisper access with a local file lock to avoid CPU-heavy concurrent Whisper requests.
• If Whisper fails and OPENAI_API_KEY is missing, that file fails.
• If OpenRouter cleanup fails or no OpenRouter key can be resolved from the environment / shell startup file / optional legacy dotenv fallback, the script still writes markdown using the raw transcript and records a warning in metadata and in the manifest.
• WHISPER_URL is intentionally not pulled from any legacy dotenv fallback, because those files may contain hostnames that are wrong for direct host-side agent use.
• Output file collisions are handled automatically by appending -2, -3, and so on.

Good response pattern

After running the script, give a short operational summary such as:

• how many files were processed
• how many succeeded or failed
• output directory path
• whether cleanup matched the legacy path fully or had fallbacks/skips

Example user requests that should trigger this skill

• "Take these voice notes in /path/to/meeting-notes/ and turn them into markdown transcripts."
• "Use the old Telegram transcription workflow on these .oga files, but without Telegram."
• "Batch transcribe all my interview .m4a files into markdown with cleanup."
• "I dropped three audio files in Downloads — convert them into transcript markdown files."