RoastCam

Settings

Where the jokes come from

Choosing a mode does not switch anything else off. It decides which settings get reached for first — sections this mode does not normally use are dimmed below, not disabled.

The badge over the video shows the mode that actually ran. It can differ from this one when a fallback kicks in, and that is working as intended.

Gemini cloud · hybrid

API key

Model

Roasts are written at thinking "low", which on a short roast usually means no thinking at all - about 1.5-2.5s on 3.8 Flash, under 1s on 3.5 Flash Lite.

Local models hybrid · local

Ollama host

Roast writer

An abliterated / uncensored model belongs here. The three at the top are the curated picks, with measured speed.

Observer (vision) — Local mode only

Leave on Auto if your roast writer is multimodal — one resident model is far faster than swapping two. In Hybrid, Gemini is the observer and this is ignored.

Needs a vision model, e.g. ollama pull qwen2.5vl:3b

Cloud mode reaches for these too, if Gemini declines to write the joke.

Voice every mode

Language

The roast is written and spoken in this language — not translated from English, so the jokes use the idiom the language actually has. The app's own buttons and labels stay in English.

Needs a voice model that carries the language. Turbo v2.5 and Flash v2.5 carry all ten; multilingual v2 has no Vietnamese, and the plain v2 models are English only.

Speech service

A real trade, measured on this machine with tools/tts-ab.js: ElevenLabs Turbo returns a clip in about 0.4s and costs $6.02 per 1k roasts; Azure HD takes about 1.0s and costs $2.65. So Azure is 2.3x cheaper and roughly 0.6s slower — which is most of the time-to-first-word this app was built to save. Hume Octave is the dearest per character and the only one that takes a stage direction, so the chosen mode reaches the performance instead of only the writing. Only the selected one is ever called; there is no fallback between them. ElevenLabs and Azure drop to the built-in Windows voice if they fail — Hume does not, and errors instead.

Local voice

The model carries 904 readers, picked by number rather than name. Audition them with tools/piper-casting.py — rendering is free and local, so hearing all of them costs nothing.

Needs a one-off install: node tools/piper-setup.js (~96MB). After that it runs offline and bills nothing per roast. About 350–500ms a line in the app, in the same range as ElevenLabs' ~420ms.

Kokoro voice

Runs on this machine through sherpa-onnx, the same runtime a phone build would use. About 0.6s a roast once loaded, nothing per character. Speaks English, Spanish, French, Italian, Portuguese and Chinese. Install once: node tools/sherpa-setup.js --kokoro

Reference clip

Every roast is spoken in the voice on this clip: a clean 5–15 second WAV of one person talking naturally. This is the only way to one specific voice at zero cost per roast.

Only use a voice you have written permission to use — a performer hired under a contract that covers synthetic use. The model's terms require consent, and so does the law in a growing number of places. Install: node tools/sherpa-setup.js --clone

Best used to audition a performer before hiring them. This model's terms also forbid harassing or discriminatory content, which conflicts with what a roast says — so ship on a voice trained from the performer's recordings, or on Kokoro, neither of which carries that restriction.

ElevenLabs key

Without a key it falls back to the built-in Windows voice.

Voice

Model

Azure Speech key

Region

HD voices only exist in some regions — eastus, westeurope and southeastasia are safe. A good key in the wrong region fails exactly like a bad key, so test after changing it.

Voice

Delivery variation 0.9

The only delivery control DragonHD offers — it reads emotion off the text rather than taking stage directions. Higher keeps repeated roasts from sounding identical; lower is steadier.

Hume API key

From platform.hume.ai → API keys.

Voice

Your own saved voices come first, then Hume's library. A voice is required — unlike the other two engines, nothing stands in for it. If Hume fails or no voice is picked, the roast does not get spoken and you get an error, rather than a different voice you did not choose. Add voices at platform.hume.ai; the app never creates one.

Stage direction

Timing 1.00

Delivery speed. Under 1 is the slower, more deliberate read a tag usually wants; over 1 rushes it.

Octave version

Volume

Master level — the voice and the shutter/drum roll together.

Look & feel

Camera look

The grade over the camera feed. Only visible while FX is on.

Roast length

Length is the biggest speed knob: the line is spoken as soon as it is written, so a shorter roast starts sooner.

Timing overlay

Reports where each roast spent its time, for comparing models.

No camera