Stop paying per character for AI voiceovers, let an AI agent set up a free voice actor on your computer

What you'll get

Install the open-source VoxCPM2 on your own computer and use it as a voice actor on call. Type some text and get audio back. Describe a voice in one sentence to create a brand-new one, or give it a short recording of your own voice and it will speak with that voice. Everything runs on your machine, so you don't pay per character. The code and models are free to use, including in commercial projects. This article includes 3 prompts you can copy and give straight to your AI agent.

Let your AI agent handle the installation and troubleshooting. You just write the script and pick the voice. An NVIDIA graphics card works best. A Mac can run it too, just more slowly.

Step 1: Listen online first

You don't need to install anything yet. First, check whether the voice sounds like what you want. The VoxCPM homepage has a row of samples you can play right away, including newsreader-style narration, game characters, angry tones, and podcast styles. To type your own text, use the official online demo hosted on Hugging Face.

It supports 30 languages. For Chinese, it also handles several dialects, such as Sichuanese and Cantonese. You don't need to tell it which language you're using. Just type.

Step 2: Let your AI agent install it and open the web version

VoxCPM2 includes a web interface. Once it's installed, you can type text and click buttons in your browser, with no code required. Hand the installation to your AI agent:

I want to make voiceovers on this computer with the open-source VoxCPM2 (https://github.com/OpenBMB/VoxCPM). Please help me:
1. Read the README first, then check my computer. Is there an NVIDIA graphics card, and how much video memory (VRAM) does it have? If there isn't one, is it an Apple silicon Mac? Also check whether my Python version is between 3.10 and 3.12. Tell me in plain language whether it will run, and roughly how fast it will be.
2. List everything you plan to install, and wait for my approval before installing anything. The model downloads automatically the first time, so tell me roughly how big it is before the download starts.
3. After it's installed, start the web interface it ships with. Tell me which address to open in my browser, and how to shut it down when I'm done.
4. Only run it locally. Do not expose the port to the internet.

A wrong graphics card or Python version often fails halfway through the install, and that wastes the most time. Checking these first avoids most of the hassle. The web interface is for you alone. If you open it to the internet, anyone can use your computer to generate voices.

Step 3: Create a voice from one sentence

In the web interface, put the voice you want in parentheses at the very start of your sentence. It then generates a brand-new voice from your description, with no recording needed. The more specific you are, the better. You can describe gender, age, tone, mood, and pace.

You could write something like this:

(A woman around thirty, warm and steady, medium pace, like telling a friend a story)Today let's talk about how to make a complicated idea simple.

Generate the same description a few times and keep the take you like best. Once a description works, save it. Use the same description for every episode, and your voice stays consistent.

Step 4: Speak with your own voice

Upload a clear recording of your own voice. The model learns your voice and can then read any new script in it. You can also add a note in parentheses to change the delivery, such as "(a bit faster, happier)".

For a closer match, add a transcript of the recording along with the audio. The model will then pick up your tone, pauses, and speaking habits too.

Record in a quiet room with only one person speaking and no background music. A phone recording made in a quiet room works fine.

Step 5: Voice a whole script in one batch

For a single line, clicking in the web interface is enough. For a full video script or podcast script, let your AI agent do the whole batch:

Here's why the steps are set up this way:

• Listen to the first paragraph first: If you don't like the voice, you only lose one paragraph instead of waiting for the whole script to finish.
• Redo only the paragraph that went wrong: Numbers and English abbreviations are the easiest to misread, and so are words with more than one pronunciation. Redoing one paragraph is much faster than starting over.
• Keep the same voice: Use the same description or the same recording for the whole script, so it sounds like one person.

Notes

• Only clone your own voice, or the voice of someone who has clearly agreed to it. The official project explicitly bans using it to impersonate people, commit fraud, or spread false claims. Label published audio as AI-generated voiceover.
• The first run downloads the model, which takes time and disk space. According to the official figures, an NVIDIA card needs about 8GB of VRAM. A Mac without an NVIDIA card can still run it, but noticeably more slowly.
• The code and model use the Apache-2.0 license, so you can use them in commercial projects for free.
• Long scripts tend to drift as they go on. Generating paragraph by paragraph and joining the pieces afterward is more stable.

References

• GitHub: OpenBMB/VoxCPM (Apache-2.0)
• Website: voxcpm.com
• Online demo: VoxCPM Demo on Hugging Face

Want your AI agent to do the voiceover for you, following clear rules? Hand it the VoxCPM skill. Just want to see what it can do first? Read the example Run AI voice synthesis free on your own computer.