bumblebee

Bumblebee Skill

A skill that automatically assembles a fragmovie-style video from an arbitrary phrase.

When to activate

• The user wants to "splice a phrase out of real movies"
• The user wants a video where other actors speak their text
• Requests like "make a fragmovie", "stitch my line out of cinema", "bumblebee"

What it does

1. Takes English text (one or more phrases as separate args)
2. Splits each phrase by sentence terminators (.!?), processes sentences independently
3. Greedy longest-match: for each sentence, finds the largest contiguous chunks of words that exist on yarn.co
4. Downloads candidate mp4s, transcribes them locally with faster-whisper (word-level timestamps), checks for exact match
5. Cuts with FFmpeg to millisecond precision, concatenates with short fade-in/out at splices and breathing pauses between sentences
6. Optional: generates N distinct variants without clip reuse via --variants N

Dependencies

• Python 3.9+ with curl_cffi, faster-whisper
• FFmpeg on PATH (or FFMPEG_BIN env var)
• No API keys required

Run

python bumblebee.py "I am your father" -o father.mp4
python bumblebee.py "Sentient is the best" -o sentient.mp4 --variants 5
python bumblebee.py "long phrase here" --variants 5

playphrase.me is consulted automatically the first time yarn fails to cover a chunk — no flag needed. The browser bootstrap (~10-15s, one-time per run) only happens if yarn actually misses; phrases yarn covers fully never touch playwright. Pass --no-playphrase to stay yarn-only (e.g. when playwright isn't installed).

Handling unreachable words (intelligent synonym substitution)

After every run bumblebee prints a machine-parseable summary line:

BUMBLEBEE_SUMMARY: {"variants_built": 5, "files": [...], "skipped_words": ["fragmovie"]}

If skipped_words is non-empty, those words were not found in yarn, in playphrase, nor in the local cache — they have never been spoken in any indexed movie or TV show, so no amount of retrying or extra sources will recover them. Do not call any TTS or generate fake audio. Instead:

1. Read the original phrase the user gave.
2. For each unreachable word, pick a contextually-appropriate synonym that

preserves the meaning of the surrounding sentence. Use the surrounding words and the user's apparent intent as context — for example:

• fragmovie → supercut / montage / compilation
• subreddit → forum / community
• blockchain → network / ledger (depending on the framing)
• Briefly tell the user which words were unreachable and which synonyms

you chose, so they can override if your choice changes the meaning.

1. Re-run bumblebee with the substituted phrase, keeping the same --variants

value as the original run.

1. If the second run still has skipped words, repeat the substitution loop

on those — but stop after at most two retries to avoid drifting too far from the user's original phrase.

This intelligent substitution loop is the correct way to handle missing words. The skill itself never invents audio; the orchestrator (you) decides when to substitute and what to substitute with.

Architecture

• bumblebee.py — CLI
• src/yarn_search.py — curl_cffi (impersonate=chrome) against getyarn.io, bypasses Cloudflare
• src/downloader.py — mp4 download via curl_cffi (yarn.co requires a Chrome TLS fingerprint)
• src/transcriber.py — faster-whisper local inference + word-timestamps + JSON cache
• src/word_matcher.py — exact match with apostrophe / prefix fuzzing
• src/phrase_splitter.py — greedy longest-match, lazy playphrase fallback on yarn miss, optional shuffling and clip exclusion for variants
• src/playphrase_search.py — Playwright-based playphrase.me adapter (lazy-init)
• src/cutter.py — FFmpeg cut with audio fade at splice points
• src/concat.py — concat demuxer

Known limitations

• yarn.co indexes English-language media only
• Whisper sometimes folds short tokens ("I", "a", "my") into longer words, so isolated short words frequently get skipped — this is an inherent limitation of word-level recognition
• A handful of "popular" yarn clips (e.g. 4e5bfded) appear in many search results and tend to be picked first; mix mode (--variants) is the workaround
• Word order is strict: can we does not match we can (swap-fuzzy is in the TODO list)