cxavireh/hotz-skills/transcribe

Audio Transcription

Use the existing Volcengine flash ASR integration for the user's requested recording. Python 3.9+ and its standard library are sufficient; no pip install or virtual environment is needed. Sending a recording to ASR is part of an authorized transcription task; a setup/check request alone does not authorize uploading audio.

Configuration

python3 ~/.codex/skills/transcribe/scripts/transcribe.py --check

This reports whether VOLCENGINE_ASR_API_KEY is available, never its value, and makes no network request. If unavailable, explain that the user must configure the key through their existing environment/secret mechanism. Do not print the environment, request the key in chat, put it in command arguments, or promise trial quotas. The service also needs the volc.bigasr.auc_turbo entitlement; --check cannot verify account permissions or balance.

Transcribe and save

python3 ~/.codex/skills/transcribe/scripts/transcribe.py '/path/recording.wav' -o '/path/transcript.md'

This saves transcript.md plus the unmodified response body in transcript.raw.json. Keep these source artifacts before any summarizing or editing. With no -o, local audio gets adjacent .transcript.md and .transcript.raw.json files; URL inputs get timestamped files in the working directory. Existing output files are preserved unless --overwrite is supplied. --json prints JSON and makes -o the raw JSON path, while still saving a readable .md companion.

The script validates HTTP and ASR business status, preserves error response bodies locally, and avoids automatic POST retries. If a request times out, explain the uncertain outcome before making another potentially billable request.

Deliver the saved transcript and any requested summary. Preserve uncertain words and speaker IDs without guessing identities. Ask about speaker names only when needed for the requested output; do not delay an otherwise complete transcript for optional labels.

Input limits

The current flash API documents a two-hour duration limit and a 100 MB file limit, with WAV, MP3, and OGG/Opus support. The script checks local size; use available audio metadata tools for duration when needed. For M4A/AAC or an unsupported container, convert a copy to a documented format using an available tool such as ffmpeg, preserving the source. Do not assume the older skill's longer format/language list or free-hour claim still applies.

API/status reference, checked 2026-09-20: Volcengine flash recognition API.