AI Voicechanger

This project is a Python scaffold for a live voice changer that uses RVC-style .pth model files and .index feature-index files.

Important: RVC .pth + .index models are voice conversion models. They convert one audio signal into another voice. They do not accept raw text directly. The practical low-latency flow is:

microphone audio -> optional speech-to-text captions/logging -> RVC voice conversion -> virtual audio device -> Discord/game

If you truly want microphone -> speech-to-text -> generated voice, that is a speech-to-text plus text-to-speech app, and the .pth/.index RVC files are not the part that reads the text. This scaffold keeps speech-to-text optional and routes the original microphone audio through an RVC backend.

What is included

• A desktop UI entry point: ai-voicechanger-gui
• A command-line app entry point: ai-voicechanger
• Device listing for microphone/output selection
• A stream pipeline that captures microphone audio and writes processed audio to an output device
• Optional speech-to-text worker for captions/debugging
• Validation for .pth and .index files
• A pluggable RVC converter interface with:
  • passthrough mode for testing audio routing
  • external-rvc mode that calls a user-provided command for conversion

Recommended Discord/game routing

Install a virtual audio cable and set it as the app output, then choose that virtual cable as the microphone in Discord or your game:

• Windows: VB-CABLE, VoiceMeeter, SteelSeries Sonar, or similar
• macOS: BlackHole or Loopback
• Linux: PulseAudio/PipeWire virtual sink/source

Install

python -m venv .venv
source .venv/bin/activate
pip install -e .[stt,dev]

If you only want audio routing without transcription:

pip install -e .

Put model files in place

Place your model files somewhere on disk, for example:

models/my_voice.pth
models/added_IVF.index

Large model files should not be committed to git.

Launch the desktop UI

ai-voicechanger-gui

Use the UI to:

1. Pick your microphone input.
2. Pick the output device that Discord or your game should hear, usually a virtual audio cable.
3. Start with passthrough to confirm routing works.
4. Switch to external-rvc and browse for your .pth and .index files.
5. Enter the RVC command template that converts {input} into {output}.

The UI also shows a checklist of what you need: microphone, virtual cable, RVC model/index files, optional STT model, and headphones.

List audio devices from the CLI

ai-voicechanger --list-devices

Run in passthrough mode from the CLI first

This verifies microphone capture and virtual audio output before adding RVC inference:

ai-voicechanger \
  --input-device "Your Microphone" \
  --output-device "CABLE Input" \
  --backend passthrough

Run with .pth and .index validation

ai-voicechanger \
  --input-device "Your Microphone" \
  --output-device "CABLE Input" \
  --backend external-rvc \
  --pth models/my_voice.pth \
  --index models/added_IVF.index \
  --rvc-command "python path/to/rvc_infer_cli.py --input {input} --output {output} --model {pth} --index {index}"

external-rvc writes each audio chunk to a temporary WAV file, calls your command, and reads the converted WAV back. This is simple and backend-agnostic, but it is not the lowest-latency architecture. For production realtime use, replace ExternalRvcConverter with a direct in-process RVC inference implementation.

Enable speech-to-text captions

ai-voicechanger \
  --input-device "Your Microphone" \
  --output-device "CABLE Input" \
  --backend passthrough \
  --stt-model small.en

Speech-to-text is for captions/logging in this app. It is intentionally not fed into the RVC .pth/.index files because those models expect audio, not text.

Latency notes

For Discord or games, latency matters. Start with:

• Wired headphones to avoid feedback
• A virtual cable as the app output
• --block-ms 80 or lower if your machine can keep up
• GPU-backed RVC inference if you add a direct backend

The external command backend is best for proving the integration. A direct backend is recommended for real-time usage.