Zero-shot voice cloning using open source models, Python, and MLX on macOS

In the James Bond series ‘Diamonds Are Forever’ (1971), the villain Ernst Stavro Blofeld and his associates employ a sophisticated voice-algorithm recorder — a bulky electronic device resembling early voice-synthesis technology — to perfectly clone the voice of the reclusive billionaire Willard Whyte. By capturing and replaying Whyte’s unique voice patterns and resonance from a small aural signature tape, Blofeld impersonates the tycoon during phone calls and communications with Whyte’s employees and security personnel. This deception allows the villain to issue orders, control Whyte’s vast casino and business empire in Las Vegas, and maintain the illusion that the real Whyte remains in charge, all while the billionaire is held captive in a remote desert hideout. The gadget enables Blofeld to operate undetected, directing henchmen and advancing his larger plot involving smuggled diamonds and a space-based laser satellite.

The voice-cloning scenes in ‘Diamonds Are Forever’ embody classic Bond ingenuity, blending 1970s camp with prescient ideas about technology-driven identity theft. This technology is no longer science fiction, it can be done on your Mac today within minutes.

Let’s get it working on your Mac first! Pull up a terminal and execute the following commands. This will install everything we need; last line it to ensure everything works. FYI, the last line might take several minutes to execute.

cd ~
git clone https://github.com/JosephMRally/Databricks-AI.git
cd voice_clone_v1
python3 -m venv venv
source venv/bin/activate
pip install --upgrade pip
brew install mlx-lm
pip install mlx-lm
pip install mlx-audio
brew install ffmpeg
pip install mlx-whisper
mlx_audio.tts.generate --model mlx-community/Qwen3-TTS-12Hz-0.6B-Base-4bit --text "Hello, how are you?" --play

Did you hear “Hello, how are you?” Perfect. Let’s move on. Record your own voice with the command below and say exactly the following “Hello, this is my cloned voice … I want you to know something, i am not real, i am fake, i do not exist … nothing exists. I would be a nihilist if I did exists.”

python clone.py --record --record_seconds 10 --output_path .

Now let’s clone our voice using the following command

python clone.py --clone --ref_audio my_voice.wav --ref_text "Hello, this is my cloned voice … I want you to know something, i am not real, i am fake, i do not exist … nothing exists. I would be a nihilist if I did exists." --text "Hi Mom, can you send me so" --play

Wow. That was cool. Now let’s see how similar they sound by using three metrics, all word-independent:

Metric │ What it measures │ Weight Timbre / MFCC │ Vocal tract shape, voice texture │ 60% Pitch / F0 │ Fundamental frequency distribution │ 25% Spectral Shape │ Brightness, resonance, bandwidth │ 15%

python my_voice_comparison_tester.py --original my_voice.wav --cloned my_cloned_voice.wav