xonika9/agent-skills / skills/x9-video-transcripts
x9-video-transcripts
VerifiedUse when the user wants offline, timestamped transcripts of their own local video or audio files, from one file to a whole folder, or wants to find where something was said in recordings already transcribed this way — «расшифруй ролики», «сделай субтитры к видео», «где в роликах я говорил про…», "transcribe these videos with timestamps". Do not use for online video links, a quick summary of what a video is about, live dictation, or translation.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 5.5 KB |
| references/models.md | markdown | 2.0 KB |
| references/onboarding.json | config | 1.4 KB |
| scripts/test_transcripts.py | script | 23.0 KB |
| scripts/transcripts.py | script | 20.8 KB |
references/models.md
View on GitHub# Models
Read this when the user wants a different accuracy, speed, or language trade-off than the default.
## Default
`ggml-large-v3-turbo-q8_0.bin` (whisper.cpp, about 874 MB) with the Silero voice-activity model. The voice-activity pass keeps Whisper from inventing text in silence and music-only stretches.
## Another Whisper model
`run --model <file>` accepts any ggml Whisper model; the official files are listed in the `ggerganov/whisper.cpp` repository on Hugging Face. Download a model only after the user chooses it, into the models directory used by `check`. Larger is not automatically better: `large-v3` (tested as `q5_0`) is several times slower than turbo and, in the comparison below, also repeated a phrase and dropped speech under music.
Existing outputs stay current until their source changes, so applying a new model to already transcribed files needs `--force`.
## Engines this script does not run
GigaAM-v3 (Russian only) and NVIDIA Parakeet v3 are faster than Whisper turbo but need their own runtimes and produce different segment layouts. Treat a request for them as a separate change to the script rather than an option.
## One measured comparison
A single comparison, not a benchmark to rely on: two Russian videos (a 12-minute narrated game-lore video with many proper names and a 20-minute vlog), Apple M4 Pro, October 2026. Error rates come from 1,120 words of hand-adjudicated reference text.
| Model | Word error rate | Speed vs real time | Notes |
| --- | --- | --- | --- |
| Whisper large-v3-turbo q8_0 | 2.6% | ~21× | Most consistent across both videos |
| GigaAM-v3 e2e-ctc (CPU) | 2.8% | ~36–50× | Mixes Latin and Cyrillic in English names |
| Whisper large-v3 q5_0 | 2.9% | ~7.5× | Best proper-name spelling; repeated a phrase, dropped speech under music |
| GigaAM-v3 e2e-rnnt (CPU) | 3.8% | ~35× | |
| Parakeet-TDT-0.6b-v3 (MLX) | 5.8% | ~57× | Most precise timestamps; more grammar and name errors |
Re-measure on the user's own recordings before changing the default; accent, microphone, and vocabulary move these numbers.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |