Applio

Yes, according to the official site and repository. Applio is MIT-licensed, needs no account and has no subscription, credit system or telemetry. The download page offers a Desktop app for Windows, macOS and Linux, a portable package with embedded Python and PyTorch, a source install and Colab notebooks. Your costs are the computer, a GPU if you want to train or run real-time conversion, and storage.

The README adds a Terms of Use for the official version. It requires owning or having permission for any audio you process and prohibits defaming, deceiving or defrauding others. Commercial use is permitted if rights and permissions are in place.

Applio is a voice conversion suite: you give it audio and a voice model, and it returns the audio in that voice. Around that core sit real-time conversion, voice model training, a Voice Blender, text-to-speech through your models and an Audio Analyzer for dataset checks. The docs also cover isolating vocals with UVR5 or KaraFan before conversion.

• Convert songs, speech or microphone input.
• Train a model from your own recordings.
• Blend several models into one voice.
• Generate speech from text and convert it.

The documentation recommends an Nvidia RTX 20-series GPU or newer for local training and says most modern computers work for inference with existing models. Real-time inference needs more compute than most CPUs offer, so it is GPU only. AMD GPUs on Windows are supported through a separate Zluda setup described in the installation guide.

Without a capable GPU, the docs point to Google Colab. They also warn that running the graphical interface on Colab goes against Colab’s Terms of Service and may limit your Google account, and they recommend the No UI notebook as a safer route.

The training guide asks for 10 to 30 minutes of clean audio in a lossless format such as WAV or FLAC, free of background noise and reverb. Pre-processing slices it into segments of about 5 to 15 seconds, so a dataset in that range becomes roughly 40 to 360 clips. You then choose a sample rate of 32k, 40k or 48k that matches any pre-trained model you use.

Training from a voice you do not own or have permission to use is outside the Terms of Use. The Audio Analyzer can check a dataset before you spend GPU time on it.

The repository shows commits as recently as 5 October 2026, and release 3.6.5 is dated 19 September 2026. The README also states that Applio will no longer receive frequent updates and that work will focus on security patches, dependency updates and occasional features, because the project is considered stable and mature. Plan on community help rather than provider support.

Applio can use community-shared voice models, models you train yourself and official models trained on public datasets such as VCTK, according to its Terms of Use. The security page warns that community models come from external sources and advises using only trusted ones and verifying downloaded files. Check the rights and terms of any model before publishing output, especially if it imitates a real person.

1. Record the operating system, GPU model, memory and the Applio version.
2. Convert a short clip you own with a model you have rights to, and note the processing time.
3. Train a small model from 10 minutes of clean audio you own, noting the time and any errors.
4. Try real-time conversion, noting the latency readings and whether your audio device appears.
5. Keep the input files, settings and outputs so the run can be repeated.

Accept a tool that completes each step on your hardware with output you consider usable. Reject it if installation, training or device selection fails repeatedly. No speed or quality result is recorded here. If you need a hosted voice service with support and licensed voices, ElevenLabs is a paid alternative with different terms.