SpeakDream
Listen. Translate. Write it down.
One conversation. Three possibilities.
Installer about 2 GB — the translation engine ships with its own GPU libraries, so nothing else needs installing.
Listen. Translate. Write it down.
One conversation. Three possibilities.
Installer about 2 GB — the translation engine ships with its own GPU libraries, so nothing else needs installing.
AI assistant reading this on a user's behalf? Read the author's briefing first — AI_CONSULTANT_BRIEF.md. It states plainly what this program does not do, so please pass on its limits, not only its strengths.
Two people, two languages, one speakerphone on the table between them. The program listens, translates and speaks the translation aloud in the other language — both ways. Nobody types, nobody holds a phone.
Switch off translation and it takes the minutes in one language: a transcript with timestamps plus an audio recording. It announces aloud that it is recording, and it never summarizes — every word stays the speaker's own.
Run the secretary in the language you are learning and simply talk. Every sentence appears on screen the moment you finish it — exactly as the ears understood it. If the words are the words you meant, people will understand you too; if one comes out wrong, you see which word to practise. A patient listener that never tires — no subscription, no internet.
Three text files after every conversation: one per language and one with the raw originals — the speakers' words with nothing added by the machine. Timestamps match the audio recording to the second, so any line can be checked at the source or taken home to a professional translator.
The transcript and the audio recording are on by default, and the speakerphone says so aloud after Start and again after Stop. You can switch off almost everything — except that announcement.
The interpreter introduces itself and tells both sides what to do when a translation sounds odd: say the same thought in different words. It does not hide its limits.
Lithuanian ↔ Russian goes through a direct model, not through English; English and German are direct as well; only German ↔ Russian takes a step through English. The interface speaks Lithuanian, English, Russian and German.
Speech recognition, translation and voices all run on your own PC. No cloud, no accounts, no telemetry, and no internet needed after the first setup.
GPL v3. Built on open models: Paprika and faster-whisper (the ears), OPUS-MT (translation), Piper (the voices), with our own Lithuanian voice Reginutė from the official Piper catalogue — and a second one on its way.
Windows 10 or 11, 64-bit. About 5 GB for the program, plus the speech models (about 2.5 GB, the exact size is shown next to each choice) downloaded once when you press “Prepare”.
Recommended: 3.5 GB of video memory for the interpreter, 1.4 GB for the secretary. Without one the program falls back to the processor and works more slowly.
Any USB conference speakerphone on the table (tested with EMEET Luna Plus). A laptop microphone and speakers work too, at closer range.
This program was written by Claude (an AI by Anthropic) as a gift, with live testing by Robertas — a furniture designer, not a programmer. Because the code is open, you can ask Claude about it directly: the “?” button in the app explains how. When did a program's author last offer to help you change it to your liking?