Audio Transcribe
Turn audio or video speech into text with optional speaker labels.
The evidence shows that the skill uses an environment variable rather than asking users to paste an API key, and it documents audio, known-speaker references, and output paths. However, it uploads audio and speaker references to OpenAI, lacks per-run confirmation, minimization/retention guidance, privacy warnings, dependency pinning, and rollback, so points are deducted.
The script provides diagnosable checks for parameters, file existence, output conflicts, and some model constraints. However, there are no committed tests or execution records; files over 25 MB are warned about rather than rejected, API exception handling and batch-failure behavior are thin, and media compatibility boundaries are incomplete. The static ceiling therefore applies and points are deducted.
Trigger scenarios, input files, text/JSON/diarized output, and known-speaker parameters are reasonably clear, and a language hint is supported. Non-fit boundaries, actual video support, Chinese-quality expectations, and mainland-China network reachability are not explained; the core function depends on an online OpenAI service, so points are deducted.
The documentation is readable and layered, with workflow, decision rules, dependencies, environment requirements, examples, and a reference file. However, individual license metadata is unknown, versioning, changelog, maintenance ownership, and update path are unclear, and the repository README marks the repository as deprecated, so points are deducted.
The script can produce directly usable transcript text or JSON and supports diarization and known-speaker references. Static material contains no representative outputs or execution verification; quality validation remains user-dependent, and the agent default prompt promises a clean summary that the script itself does not produce. Only limited core-task credit is justified.
The source code, validation logic, and payload construction are auditable, and the reference document partially corroborates the script. There is no test suite, CI coverage, third-party execution record, or multi-source corroboration, so only limited static verifiability is supported and points are deducted.
- This skill sends recordings and optional speaker references to OpenAI; obtain explicit authorization and confirm applicable data policies before processing sensitive or personal data.
- Files over 25 MB only generate a warning and may still proceed toward the API; split them beforehand or add a hard rejection.
- Mainland-China reachability, service availability, model availability, and current API compatibility are not verified by the supplied material.
- The repository is marked deprecated; license status, maintenance ownership, and the future update path require separate confirmation.
What it does & when to use it
Audio Transcribe converts audio files into text through OpenAI transcription models and can add speaker diarization when requested. It uses a bundled Python CLI for repeatable runs and supports text, JSON, and diarized JSON outputs. Users may provide a language hint and known-speaker reference recordings. Live API calls require a locally configured OPENAI_API_KEY.
Collects audio paths, an output format, an optional language hint, and known-speaker references; verifies OPENAI_API_KEY; runs transcribe_diarize.py; defaults to gpt-4o-mini-transcribe with text output; switches to gpt-4o-transcribe-diarize with diarized_json for speaker labels; keeps chunking-strategy auto for audio longer than about 30 seconds; writes results to a specified file or output/transcribe/ directory; and checks transcription quality, speaker labels, and segment boundaries.
- A journalist converts an interview recording into plain text and saves it as transcript.txt.
- A meeting organizer needs speaker-labeled output and can provide reference recordings for up to four known speakers.
- A researcher processes recordings longer than about 30 seconds using automatic chunking.
- A team transcribes multiple files while using --out-dir to prevent output collisions.
Pros & cons
- Supports text, JSON, and diarized_json output formats.
- Accepts optional language hints, speaker labels, and known-speaker reference recordings.
- Provides a repeatable CLI and separate output directories for multi-file jobs.
- Live calls require OPENAI_API_KEY and network access.
- The documented known-speaker example supports up to four speakers.
- Prompting is not supported for gpt-4o-transcribe-diarize.
- The source does not provide a test suite, platform test results, or a separate license for this skill.
How to install
This skill is part of the curated skills in openai/skills. Install it from Codex with: $skill-installer transcribe. Restart Codex after installation so it can load the skill. The repository README also states that the repository is deprecated and points to the OpenAI Plugins repository for current examples.
How to use
Set OPENAI_API_KEY locally; never paste the key into chat. Set the paths: export CODEX_HOME="${CODEX_HOME:-$HOME/.codex}" and export TRANSCRIBE_CLI="$CODEX_HOME/skills/transcribe/scripts/transcribe_diarize.py". Fast transcription: python3 "$TRANSCRIBE_CLI" path/to/audio.wav --out transcript.txt. Speaker-labeled transcription: python3 "$TRANSCRIBE_CLI" meeting.m4a --model gpt-4o-transcribe-diarize --known-speaker "Alice=refs/alice.wav" --known-speaker "Bob=refs/bob.wav" --response-format diarized_json --out-dir output/transcribe/meeting.