Speech to Text

Neu in Version 21.04.0.

Warnung

Speech to text does not work with version 21.04.2 due to Vosk API issues. Use version 21.04.1 or 21.04.3 and later versions.

Install Python

Python 3 needs to be installed on your computer (details see below for Linux and Windows). Once Python is installed, follow these steps to put Python into a virtual environment (afterwards Python is copied to the venv folder)

De-install Python

To remove the installed venv package got to Settings ‣ Kdenlive Settings ‣ Environment ‣ Python and Delete venv.

It will completely remove the venv folder with all installed packages. Note that this does not remove the downloaded models (vosk/whisper) that can still take quite some HD space

Linux

On most Linux distributions python is installed by default. You can check if that is the case for you too by running python3 -V in a terminal. If python is missing just search the internet, there are lots of instructions around.

Windows

Download python from https://www.python.org/downloads/ for installation on your computer.

Speech Engines

To install the speech engines go to Settings ‣ Configure Kdenlive ‣ Speech to Text.

VOSK

Vosk download dependencies

Vosk is not installed

When you switch to VOSK for the first time you have to install the missing dependencies first.

Path where VOSK is installed:

  • Linux: ~/.local/share/kdenlive/venv/Lib

  • Windows: %LocalAppData%\kdenlive\venv\Lib

If you have installed VOSK in an earlier Kdenlive version already and now you have chosen the venv folder for Python, you can delete the past installed VOSK libraries by using following command in a console: pip uninstall vosk srt

Install a Language

Goto Settings ‣ Configure Kdenlive… ‣ Speech to Text page and select the speech engine VOSK.

Click on the link to get a language model.

download link

Drag & drop the language you want from the vosk-model download page to the model window, and it will download and extract it for you.

download model

If you have problems or check for updates click on the Check configuration button.

The VOSK speech models are stored here:

Linux: ~/.local/share/kdenlive/speechmodels

Windows: %AppData%\kdenlive\speechmodels

Whisper

Neu in Version 23.04.

OpenAI-Whisper is a speech recognition model for general use. It is trained on a large dataset of diverse audio and is capable of performing speech translation, and language identification.

Whisper is slower than VOSK on CPU, but it is more accurate than VOSK. Whisper creates sentences with punctuation marks, even in Base mode.

Whisper download dependencies

Whisper is not installed

When you switch to Whisper for the first time you have to install the missing dependencies first (about 2GB to download).

Whisper installed

When all is correct configured, you get this screen.

Path where Whisper is installed:

  • Linux: ~/.local/share/kdenlive/venv/Lib

  • Windows: %LocalAppData%\kdenlive\venv\Lib

The Whisper speech models are stored here:

Linux: ~/.local/share/kdenlive/opencvmodels

Windows: %AppData%\kdenlive\opencvmodels

Model Select the model. More details on the Whisper source code page (default: Base) .

Language Select the language if Autodetect is not accurate (default: Autodetect)

Device For compatibility purposes only CPU is available

Translate text to english This translates non-English text to English during recognition

You can check for updates by clicking on Check configuration

If you have installed Whisper in an earlier Kdenlive version already and now you have chosen the venv folder for Python, you can delete the past installed Whisper libraries by using following command in a console: pip uninstall openai-whisper

Speech recognition

Select the speech engine

Neu in Version 23.04.

Enable Menu ‣ View ‣ Speech Editor menu item.

change the speech engine

Click on the Hamburger Menu application-menu and select Configure Speech Recognition. This brings you to Configure Speech to Text, select the engine and click OK.

Translate to english is only available with the Whisper speech engine. It translates non-English text to English during recognition.

Creating subtitle by speech recognition

Speech to text subtitle

Shown with the VOSK engine

  1. Mark the timeline zone you want to recognize (adjust the blue line)

  2. Click on the Speech recognition icon

  3. Choose the language

  4. Choose how the selected zone should be applied

  5. Press on the Process button

The subtitle gets created and inserted automatically.

Bemerkung

Only timeline zone is implemented for now in automatic subtitles.

Remark to 4: The default is to analyze only the Timeline zone (all tracks) (the blue bar in the timeline ruler). Set the zone in the timeline to what you want to analyze (use I and O to set in and out points). Selected clips option analyses the selected clip only.

Creating clips by speech recognition

This is useful for interviews and other speech-related footage. Enable the Menu ‣ View ‣ Speech Editor menu item.

Speech editor

Shown with the VOSK engine and search enabled

Select a clip in the Project Bin.

  1. If needed set in/out point in the clip monitor and enable Selected zone only selection box. This will only recognize the text inside the zone.

  2. Choose the correct language when the VOSK engine is selected. Or choose the Whisper engine by click on Configure Speech Recognition (see configure speech to text)

  3. Press the Start Recognition button.

  4. Select the text you want. Holding CTRL or Shift to select several texts.

  1. Choose: Create new sequence with edit creates a new sequence with each timecode-text as a single clip, or Insert selection in timeline at playhead postion, or to Save edited text in a playlist file which appears in the project bin.

  1. Zoom in or Zoom out of the text. Remove non spech zones deletes all „No speech“ entries at once.

  2. Add a Bookmark. You can jump to these bookmarks in the timeline with the Alt + arrow shortcut or edit the bookmark by double click.

  3. Delete the selected text.

  4. Here you can search in the text.

  5. And navigate up or down in the text.

Silence detection

This works with the VOSK engine only.

Open the clip in the clip monitor and open the speech editor window (Menu ‣ View ‣ Speech Editor) .

Select your language or Speech Engines and download the model for it.

Then click Start Recognition button.

Once this is done, choose under point 6 from above to Remove non speech zones at once. Or click on the time-code where „No speech“ is indicated (hold CTRl to select several items at once) and just hit the Delete key.

Repeat the operation for all the parts you want to remove, including where someone says what you do not want to include in your final edit.

Once finished, make sure Selected zone only is disabled, click on the Save edited text in a playlist file button (above under point 5) and after few seconds a new playlist is added in the Project Bin without silence and without the text you do not want.