Use cases
The same program serves a call, a lecture and a folder of recordings, but the setup differs. Pick the situation that is yours: each section says what happens, what you need for it, and which guide walks you through it.
- A call in both directions
- A call, incoming only
- Lectures, video and recordings
- Practice and demonstrations
- Without a graphics card
- A subtitle backlog
- Translating subtitles and documents
- Several workstations, one machine with a graphics card
A call in both directions
The other party's words appear as captions and a translation over the call window, and the interpreter says them into your headphones. What you say goes out to the other party in their language, in your voice, through a virtual microphone that you select in the call application.
Two virtual cables do this, with fixed roles: one brings the call in, the other carries the interpreter out. The Usage tab of the settings window writes out, with the actual device names, what to set where in the call application, and then reads back what the application really uses and compares the two.
The program recognizes ten call applications: Microsoft Teams (new and classic), Zoom, Webex, Discord, Slack, Skype, Signal, WhatsApp and Viber. Two things to know: the call audio in your ear lags about 0.2 seconds behind the direct path, and without the program running the cable is silent, so the secondary ringer belongs on the headphones.
What you need: a graphics card here or on another machine on your network, two virtual cables (VB-CABLE and the A+B pack), headphones
A call, incoming only
If you only want to understand the other party, the outgoing direction stays off and one cable is enough: the mode "Call – incoming direction only" on the Usage tab sets it up. Captions, translation and the interpreter in your headphones work the same way.
With many programs even the cable is unnecessary: the default audio source is the speaker's own output. Teams is the measured exception – no loopback capture heard its audio on the test machine, and that is what the cable is for.
What you need: a graphics card here or on another machine, headphones; one virtual cable if the call application's audio cannot be captured
Lectures, video and recordings
Whatever the computer plays – a lecture stream, a video, a webinar – gets captions and a translation in the transparent band. The source is "the computer's audio"; for lectures the "Accurate" profile suits best, with longer sentences and more context for the translator.
Every session is recorded in its final form, so the transcript is complete even with translation switched off. The Recordings window searches the text and exports it as bilingual text, as SRT subtitles (the original or the translation) or as a CSV table. The recordings are files on your own disk, kept for 30 days by default.
What you need: a graphics card here or on another machine; headphones only if the interpreter speaks
Practice and demonstrations
With the microphone as the source, your own speech is captioned and translated, and with the interpreter on you hear yourself in the other language. It works only with headphones: from a speaker the interpreter's voice would go back into the microphone, and the program would start translating its own translation – the program asks about this once per run.
What you need: a graphics card here or on another machine, headphones
Without a graphics card
The captions can come from Windows Live Captions instead of our caption server: the program starts the system's recognizer, hides its window and turns its text into sentences. The interpreter can speak with the in-process voice, which runs on the CPU.
What you give up: there is no model choice (the quality is that of the Windows language pack), the segmentation is Windows's own (its controls are disabled), and the language must be set exactly in advance – no candidate list, no automatic detection. Translation still needs the translator service, typically on another machine with a graphics card.
What you need: Windows with the language pack of the speech; the translator on another machine for translation
A subtitle backlog
A folder of recordings becomes subtitle files from the command line, without the caption window. The batch script starts a separate client per file, several at once – by default as many as the server's pool allows –, skips the files that already have a newer subtitle next to them, and sums up at the end how many succeeded.
The subtitle is named with the language, so the Hungarian one is eloadas.hu.srt next to the video; the whole mode sends each file in one request for the best quality. Nothing is uploaded anywhere: the audio goes to your own caption server and comes back as text.
What you need: the caption server on a machine with a graphics card, FFmpeg (the client downloads it on request)
Translating subtitles and documents
Three command-line clients translate files with the same local translator: subtitles (.srt, .vtt, .ass/.ssa) with the timing untouched, Markdown documents with code, tables and formatting untouched, and string tables (.resx, PowerShell messages.psd1) with the keys and comments untouched.
Only the translatable text is sent; everything else is written back byte for byte. A check mode runs without the server, a folder input walks the whole tree with a per-file lock, and the access key is given as the name of an environment variable, never on the command line.
What you need: the translator on a machine with a graphics card
Guide: Translating subtitle files, markdown and resource files
Several workstations, one machine with a graphics card
The four parts can run on separate machines: the caption window on each workstation, the caption server, the translator and the voice service on the machine that has the graphics card. The workstation then needs no graphics card at all; the program only needs the services' addresses and a key for each on the Servers tab.
The network costs little: on a wired local network the measured overhead is a few milliseconds, and the bandwidth is small – 16 kHz mono upstream, 24 kHz mono downstream, the latter only while the interpreter speaks. The three engines together need the video memory in the table on the How it works page, which is why a single card is often better shared this way.
What you need: one machine with a graphics card on the network, keys for each workstation from the account service
Not sure which setup fits you?
Tell us how you would use it and how many people; we answer with a setup and a quote.