PolyCaption

Real-time multi-language conversation transcriber. Captures audio, transcribes each phrase with faster-whisper, translates it locally with argos-translate, and shows every language you configure in an overlay that is invisible to screen capture / screen share.

Pick any languages argos supports via the languages config key — e.g. ["de", "ja"] for German + Japanese, or the default ["en", "es"]. Whoever speaks one of those languages gets the others filled in. Translation pivots through English when there's no direct package (so de↔ja works without a de↔ja model).

The full list of installable language packages is the argospm index. Almost every package is X ↔ English, so the rule of thumb is simple: any language with an English package in that index can be combined with any other.

How it works

• Speaker audio is captured via VB-Cable's CABLE Output device (set your call/app output to CABLE Input).
• Your microphone is captured too (set capture_mic: false to disable).
• Each phrase is transcribed in the spoken language, then every other configured language is filled in by argos. All lines render as blue text, updating live as you speak.

Setup

1. Install VB-Cable and reboot.
2. pip install -r requirements.txt
3. Copy config.example.json to config.json and adjust it to taste (see Config below). config.json is gitignored, so your local tweaks (e.g. audio_device_index) won't get committed. If you skip this step, main.py creates a default config.json for you on first run.
4. python main.py
   • First run will download the Whisper small model and the EN↔ES argos packages automatically, so it will take a bit longer than usual.

Hotkeys

Keys	Action
Ctrl+Shift+↑/↓	Scroll transcript history
Ctrl+Shift+←/→	Move (tap) / resize width (hold)
Ctrl+Shift+ -/=	Opacity down / up
Ctrl+Shift+R	Reset / clear all text
Ctrl+Shift+Q	Quit

Config (config.json)

There's no need to create this file by hand — config.example.json has every key with its default value, and main.py will write a fresh config.json from those defaults on first run if one doesn't exist yet. Edit the values below to taste.

Key	Default	Notes
languages	["en", "es"]	ordered list of argos codes to transcribe & display; order = display order
whisper_model	small	base faster / medium more accurate
partial_model	null	a faster model (e.g. tiny/base) for the live partial line, so it feels instant while the final text still uses whisper_model. null reuses whisper_model
speaker_language	null	pin the speaker's language code (e.g. ja) to skip detection — faster & more accurate. null = auto-detect
mic_language	null	pin your language code the same way
whisper_device	null	cuda / cpu; null auto-tries CUDA then falls back to CPU
whisper_compute_type	null	override e.g. int8 / float16; null picks a sensible default per device
capture_mic	true	also transcribe your own mic
energy_threshold	100	RMS level above which audio counts as speech. Raise this (e.g. 300+) if the mic hears your speakers and phrases never end (no pause is ever detected)
speaker_max_seconds	30	hard cap; a phrase with no detected pause is force-split here
speaker_pause_threshold	0.8	seconds of speaker silence that ends a phrase
mic_pause_threshold	0.6	seconds of mic silence that ends a phrase
mic_max_seconds	25	hard cap on a single mic phrase before it's force-split
mic_resume_delay	0.8	seconds to ignore the mic after the speaker stops (prevents echo on speakers). Set 0 if you use headphones
live_transcription	true	show an updating partial line while you're still speaking
partial_interval	1.2	seconds between live partial-transcript updates
commit_after_seconds	8	for long phrases, transcribe in chunks of this length instead of waiting for silence
beam_size	1	greedy = lowest latency; raise for accuracy
audio_device_index	null	fallback device index if CABLE Output isn't found