An open Codex Skill that turns a user-provided local audio file into a faithful transcript. It is published by HA7CH but has no dependency on ANC, a HA7CH vault, or a specific workspace layout.
Supported direct inputs include WAV, MP3, M4A/MP4 AAC, FLAC, Ogg Opus, and WebM Opus. Long recordings are split locally into ten-minute WAV chunks before transcription. WeChat SILK works when a compatible rust-silk decoder is installed.
Ask Codex:
Install the Skill from https://github.com/HA7CH/anc-transcribe-audio
Then install its runtime dependency:
brew tap wangnov/tap
brew install codex-asrffmpeg is also required when a recording is longer than ten minutes:
brew install ffmpegAfter installation, ask Codex to use $anc-transcribe-audio on a local audio file. Include the upload authorization in the same request so Codex does not need to ask again:
Use $anc-transcribe-audio to create one complete Markdown transcript. I authorize uploading this specific audio to the ChatGPT transcription endpoint for that purpose.
The Skill transcribes every part before atomically creating a single Markdown file. A failed part cannot be mistaken for a completed transcript.
The Skill invokes the open-source codex-asr CLI. That tool reads the ChatGPT login already stored by Codex Desktop and uploads the selected audio to ChatGPT's non-public backend-api/transcribe endpoint. The Skill does not scan directories, persist transcripts, or copy authentication data.
This is an unofficial local adapter, not the official OpenAI Audio API. The upstream endpoint has no public stability or service-level guarantee and may change without notice. Use it for personal local workflows, not as a shared production transcription service.
bash scripts/test-transcribe-audio.sh
python3 /path/to/skill-creator/scripts/quick_validate.py .MIT. The runtime dependency codex-asr is a separate MIT-licensed project by its own contributors.