Cloning a voice takes 14 seconds of audio on Fish Audio and an API key. With that, Claude Code answers me out loud in my own voice instead of Alexa’s, and it also works on my computer’s files. The full setup took me one afternoon and €0, and the video’s opening, where my voice says “I hear you perfectly”, was generated with that cloned voice without me recording anything.
I’ve spent 3 years talking to Alexa and the conversations always end the same way: “I didn’t understand that”, “that skill isn’t available”, “sorry, I don’t know how to do that”. Meanwhile, right next to the speaker, I have an entire computer that does know how to do all of it. So I made the logical switch: an assistant that lives on my machine, touches my files, and answers me out loud.
What exactly I built
A voice assistant needs two pieces: someone who thinks and a voice that sounds, and here both run on the same computer. The thinker is Claude Code, which works with my files and runs whatever I ask it; the sound is my cloned voice on Fish Audio, which plays out loud when the model finishes writing its answer. With that, you talk to it and it acts on the machine without touching the keyboard.
The difference from Alexa is the brain: since it’s an agent with access to my machine, it can read a file, edit code, or chain tasks while we talk. Alexa was a speaker with a dictionary; this gets a bit closer to the movie Her, but set up by me and on my own computer. If what you want is a 24/7 assistant on a server, I already told you how to install OpenClaw safely on a free server; the fun part here was something else: making it sound like me. If the goal is controlling the house and not the computer, Home Assistant documents its own voice control.
Cloning my voice in Fish Audio step by step
Fish Audio gives you three paths: instant voice with a short audio, designing a voice from a written description, and professional voice cloning, which asks for more minutes of recording but delivers the result you hear at the start of the video. For the demo I used the instant one and 14 seconds of audio were enough.
- Go to fish.audio and sign up.
- Hit “create voice” and choose “instant voice”.
- Record the text it gives you or upload an audio you already have. I had to read “Last night I looked out the window and the sky was full of stars…”.
- Hit analyze, give it a name, and keep it private. Watch out here: it’s your voice, so don’t make it public.
- Write any text, hit generate speech, and listen to the result.
You can add emotion tags to the text: angry, happy, laughing. I put “laughing” in front of a line from one of my video scripts and the laugh comes out very natural. It’s pretty creepy. There’s also a preview model that sounds even better, though they’re still working on it.
If listening to your own voice gives you the creeps, design mode makes voices from scratch: I described “voice of a sweet 20-year-old girl” and it generated several samples to choose from. The more detail you put in the description, the better it comes out, though that one has a slight robotic touch.
Connecting the voice to Claude Code with the MCP
The quick connection is the Fish Audio MCP: you add it to Claude Code, authorize in the browser, and it answers you out loud. Three steps, and none of them require writing code.
- Add the Fish Audio MCP to Claude Code from the terminal.
- When you open Claude, it warns you there’s an unauthenticated MCP. Run
/mcp, choose Fish Audio, and authorize in the tab that opens in the browser. - Just ask for it as is: “I want you to answer me using Edu’s voice from Fish Audio”.
With that it already works as an assistant: you write to it, it processes whatever it is, and it plays the answer with the voice. The first test came back spoken, and that was the end of what had to end: Alexa out of the house.
From the MCP to the API: free until the end of November
The MCP does the job, but there’s a money detail that changes the decision: the Fish Audio API is free until the end of November. Since it’s free, better to have the answers come from calling the API directly.
- On Fish Audio, go to the developer section, create an API key, and copy it without showing it.
- Save it in your environment variables file (
~/.zshrcon my Mac; same on Linux and slightly different on Windows):
export FISH_AUDIO_API_KEY="tu_clave"
- Reload the terminal configuration:
source ~/.zshrc
- Tell Claude to set the MCP aside and implement the API following its official documentation. I passed it the link to the docs and it connected on its own: it read the documentation, grabbed the key from the environment, and set up the hook that plays every answer out loud. I didn’t write a single line of integration.
To talk to it by audio myself, I use the voice input in the Warp terminal: I hit the mic, I speak, and it answers me with my cloned voice. If Warp isn’t your thing, I already showed you how to make an assistant understand audios for free with the Whisper API from Groq.
What went wrong and what I’d change
Three things failed: the accent of the first audio, a typo of mine in the test script, and the voice designed by description, which sounds more robotic than the cloned one. The first two are fixed by changing the text; the third is a matter of choosing the cloning mode.
The first thing that failed was the accent. My first audio came out in Argentine Spanish because I have a skill installed from Engram, a project by an Argentine guy, and the model caught his way of writing. It’s fixed in one message: “speak in Spanish from Spain”. But it’s a useful warning: everything you have installed slips into how the answer sounds.
The second one I brought on myself. I pasted the script of a video to test the voice, and the “a a new” that came out in the audio was written that way by me in the script. The voice reads exactly what you give it: if the text has an error, the error gets heard.
And third, the voice designed from text sounds a tiny bit robotic next to the cloned one. To imitate a real voice, cloning (the instant one already convinces, and the professional one even more). To invent new voices, design mode, knowing it shows.
What I’d change: removing the terminal from the middle. Right now the conversation happens inside Claude Code with Warp in front, and what I want is to send it an audio over WhatsApp and have a real-time conversation, without opening anything. The truth is I already like the result; it actually sounds better than my own voice to me, because listening to myself has always made me a bit uncomfortable.
The next step: a Jarvis with WhatsApp
The next step is an agent that receives audios over WhatsApp, talks in real time, and works without you sitting in front of the computer. The voice and the brain are already tested; what’s left is removing the terminal from the middle.
I want to release it open source so you can set it up yourselves, so if you’re interested in that series of tutorials, let me know in the comments — I read them.
In the meantime you can clone your voice today: sign up at Fish Audio, 14 seconds of recording, and an API key in the environment file. And if you want the agent to already work on its own while you sleep, the optimal OpenClaw configuration gives you the autonomy base to plug the voice into.



What do you think?
Leave your opinion, question or suggestion. Comments are synced with GitHub Discussions .