RVC Voice Conversion Plugin

February 8, 2025 ยท View on GitHub

The RVC Plugin allows to replace one voice with another. This means the input is an existing audio.

(It is not Text-to-Speech which works with Text as Input and generates audio from this.)

You need a RVCv2 Model for this to work.

Configuration

  • Under the General-Tab, select the RVCv2 Model (and Index) file you want the voice to be converted into.

    • As f0method select rmvpe for best quality, or crepe (or others) for faster but lower quality.
    • Set f0up_key to a value that fits the conversion pitch. This is depending on the Input and Output Voice.
      • lower (~ -12) if the voice conversion is female to male
      • higher (~ +12) if the voice conversion is male to female

  • Under the Audio conversion-Tab, select the source from where you want to convert the voice from in voice_change_source.

    • Own Voice works by using Whispering Tigers audio recording logic. So same VAD, Speech Pause and Timelimit settings apply as when using Speech-to-Text models.

      You also need to enable Speech-to-Text Enabled and Automatic Text-to-Speech Checkboxes in the Speech-to-Text Tab of the main application.

    • Own Voice (Realtime) works by starting a seperate thread. All Realtime settings can be found under the Plugin Realtime-Tab.

    • Set it to Text-to-Speech when using any TTS Model/Plugin to convert the TTS Voice into the RVC Models voice.

  • Under the Model-Tab,

    select the RVCv2 Model (and Index) file you want the voice to be converted into.

    • device Select your GPU for best performance. If you have one GPU, select cuda:0 if you have an NVIDIA GPU, or direct-ml:0 or direct-ml:1 for AMD/Intel.

      (Try to change this if it is too slow and does not seem to use the correct GPU, or you get the error 'RVC' object has no attribute 'tgt_sr')

    • half_precision Enable this when using a GPU and to increase speed for slightly less quality and also less memory usage.

  • Under the Realtime-Tab (Only needed in Realtime mode)

    • rt_block_time Is the audio block length. Lower values reduce the delay but increase the GPU usage and possibly lower the quality.
    • rt_crossfade_time is the time each block is faded over the next to make it more seamless.
    • rt_extra_time is added audio data which can increase quality.
    • rt_input_device_index Is the Input Audio device. (should be your microphone. [Using NVIDIA Broadcast can heavily increase quality for higher GPU time])
    • rt_output_device_index Is the Output Audio device where the audio is played to. (If you want to use it in voice-chats, you have to select a Virtual Audio Cable here and the same in voice-chat applications. Find a VB-Cable Driver here https://vb-audio.com/Cable/)
    • rt_threshold defines the threshold at which volume it records the audio. Too high and it might cut out too much or not record at all. Too low and it might pick up too much noise.

    Press the start / restart / stop Realtime Button to manually start the thread or restart when you changed some settings.

    (If you set voice_change_source from the Audio conversion-Tab to Disabled and press the start / restart / stop Realtime Button, it will stop the thread.)