Saltar al contenido principal

Dubbing

Arrives in the next build

The Dubbing window arrives in the next build. It is not in 2.2.0.

The Dubbing window takes a video file and makes the same video in another language. It transcribes the speech, translates each line, voices it and mixes the result. Each line can be read, corrected and spoken again.

Open it from the rail: right-click an empty part, Add a widget, then Dubbing. See the menu bar.

What it needs​

Dubbing runs on Windows only for now. Its voice engine, Qwen3-TTS, exists for Windows only. The window lists what is missing, with the place to fix it:

  • ffmpeg.
  • A GPU Whisper model, in Settings › Voice.
  • Qwen3-TTS, in Settings › Voice.
  • Your voice sample, when you dub in your own voice.
  • A local model, when the translation runs locally.

The Dub button stays grey until all of them are there. See voice settings.

Starting a dub​

  1. Choose a video: MP4, MOV, M4V, MKV, WebM or AVI.
  2. Into: the target language. Qwen3-TTS speaks ten: Chinese, English, German, Italian, Portuguese, Spanish, Japanese, Korean, French and Russian.
  3. Voice, one of three:
    • My voice: your recorded voice sample.
    • A library voice: a voice from your library, picked in a menu.
    • The person in the video: a clone of the speaker.
  4. Translation: local, with the model named on the button, or Cloud. The cloud model is billed for each translated line.
  5. Dub.

The person in the video asks for a tick first:

This is my voice, or I have the consent of the person who speaks in this video.

That consent covers one video. Another video asks again. The cloned voice is cut from this video and never enters your voice library.

One dub runs at a time.

The steps​

The window shows the steps in order: Sound, Transcript, Translation, Shortening, Voice, Review, Mix, Done. Shortening appears only when it ran. A progress line counts seconds of audio during the transcript, and lines after that.

Each line is voiced in the target language. A clip longer than its place is sped up, up to 1.25 times. A line still too long gets a shorter translation. The engines take the graphics card one after the other.

The dub waits while the voice assistant speaks. Cancel stops it and keeps the lines already done. Resume picks it up at the first step left to do. It also resumes a dub the app closed on.

Reading and correcting the lines​

The table has one row per line: Time, Original, Translation and Voice. Each row plays the original segment and the dubbed clip.

  • Edit a translation, then Save or Ctrl+Enter. The row says corrected by you. Its clip is erased, and the dub goes back to Review. The line is voiced again at Make the video.
  • Speak again voices one line a second time, with the same text. The dub then runs on to the end.
  • The badge in the Voice column reads fits, a speed such as ×1.10, too long, not spoken yet or not translated.

Review and export​

A line still too long stops the dub at Review. A line with no translation stops it too. The window gives the count of each. You can shorten them, or tick a box to export anyway:

  • An overlapping line runs over the next one.
  • A line with no translation keeps the original sound.

Make the video voices what is left, mixes and writes the file. A subtitle file comes with it. Open the folder shows both.

Every dub keeps its own folder, under ~/.mnemosyne/generations/dubs. The list on the left holds every dub, newest first.

For agents​

An agent can prepare a dub through the app's door dub.start. It runs the transcript and the translation, with the local model only, then stops at Review. No voice speaks from that door. The agent cannot choose the voice of the person in the video. You read the lines in this window and make the video yourself. dub.get reads a dub or the list.