Transcribe

Transcribe — Audio → Transcripts → GitHub

Use when the user gives you an audio file or link (Dropbox shared link, direct URL, uploaded file, local path) and wants it transcribed. Produces two transcripts per audio — verbatim.txt (word-for-word, fillers included) and intended.txt (clean, readable) — and commits each into its own dated folder in a private GitHub repo, then pushes.

Configuration Variables

Set these at install time (the natural-language install prompt asks for them):

Prerequisites

Procedure

1. Get the audio locally

Any of these work; pass the source string straight to the script:

• Dropbox shared link — the script auto-converts www.dropbox.com/s/... → dl.dropboxusercontent.com and appends ?dl=1 for a direct download. Works with ?dl=0 links too.
• Direct URL (http/https) — downloaded with a browser User-Agent.
• Uploaded file — Hermes desktop uploads land somewhere local; pass the path.
• Local path — pass as-is.

2. Run the transcription

"{{TRANSCRIBE_VENV_PATH}}/bin/python" "{{TRANSCRIBE_SCRIPT_PATH}}" "<SOURCE>" --out "{{TRANSCRIPTS_REPO_PATH}}" [OPTIONS]

Defaults: --model {{TRANSCRIBE_MODEL}} (CPU tradeoff), language auto-detect, both verbatim + intended.

Backend reality check (measured 2026-08-02, medium model, 10.6 s audio, 6-core EPYC CPU, no GPU):

Never use --backend transformers on a CPU-only box except as a compatibility check — it will make the user wait hours for a podcast.

Long audio: with ct2, transcription is roughly realtime (1 h audio ≈ 1 h CPU, single transcription; medium, int8). Still worth running in background with notify_on_complete=true for anything over ~15 min, and tell the user the ETA. Do NOT block the chat.

3. Verify output

The script prints OUTDIR:<path> and the word counts. Check:

ls -la "{{TRANSCRIPTS_REPO_PATH}}/<YYYY-MM-DD>_<slug>/"

Expect: verbatim.txt, intended.txt, meta.json (+ words.tsv / intended.srt if requested). Spot-check both texts — if a transcript is empty or obviously wrong, rerun with --language <code> explicitly (auto-detect occasionally fails on short clips).

4. Commit and push

cd "{{TRANSCRIPTS_REPO_PATH}}"
git add <YYYY-MM-DD>_<slug>/
git commit -m "transcribe: <slug> (<lang>, <model>, <duration>s)"
git push

Never commit the original audio — GitHub caps files at 100 MB and repos stay lean without LFS. The source is recorded in meta.json; the downloaded audio lives in the ephemeral work dir (/tmp/transcribe-work/) — if the user wants it kept, copy it to an audio archive folder before cleaning up.

5. Report

Give the user:

• The folder path (local) and the GitHub URL: {{TRANSCRIPTS_REPO_URL}}/tree/main/<folder>
• Language detected, model/backend used, audio duration vs transcription time
• A 1–2 line sample of each mode so they can see the verbatim/intended difference

Pitfalls

• Hermes profile HOME isolation — Under a Hermes profile, ~ may resolve to the profile's sandbox home (e.g. ~/.hermes/profiles/<profile>/home/), so the HF model cache and any path built from ~ land in odd places. Always use absolute paths for the venv, script, and repo (the {{VAR}} values). The model only downloads once and is reused.
• First run is slow — downloads the model (~1.5 GB for medium) from HuggingFace, and the ct2 backend converts HF→CT2 once. Subsequent runs skip both. Don't panic if the first run takes several minutes before "model loaded".
• transformers 5.x tokenizer failure — crisperwhisper 2.0.1 was built for transformers 4.x. If you see OSError: Can't load tokenizer for 'nyralabs/...', it's usually an interrupted first download (re-run fixes it). If it persists, pin transformers<5 in the venv: pip install "transformers<5".
• No GPU — on a CPU-only box, large will be slow and may OOM with word timestamps; stick to medium unless the user asks otherwise.
• ct2 float16 on CPU fails — ValueError: Requested float16 compute type... The ct2 backend defaults to compute_type="float16", which needs a GPU. The script forces compute_type="int8" when the backend is ct2; don't remove that.
• First ct2 run converts the model — HF→CT2 conversion of medium on CPU takes ~10–15 min once; cached afterwards. The transformers backend has no conversion step but is slower per-decode.
• Unauthenticated HF downloads — fine but slower; don't block on rate limits, just retry.
• Empty transcript on silence — by design (both backends decode with begin-of-sequence suppression disabled). If the audio is genuinely silent, that's the correct output.
• Dropbox links with & in the URL — quote the source argument in the shell.

Verification checklist

• verbatim.txt and intended.txt non-empty, language looks right
• meta.json has source, duration, model, backend, language
• git status clean after push; git log --oneline -1 shows the new commit
• User can open the GitHub link (private repo — their account only)