LinnetTTS: tiny speech model talks in real time on a single phone CPU core

In the browser (no installation)

The developers have an online demo on Hugging Face. You can try the model right away there without installing anything. Sometimes you have to wait a moment for the demo to wake up.

Open online demo

Windows 10 / 11

Windows 10 / 11: not officially tested. The source only refers to the ready-made Python package linnet-tts generically as a "Python wheel"; a Windows build isn't explicitly mentioned. If the install fails, use WSL2 (Linux inside Windows) with the Linux instructions, or try the online demo in the Hugging Face Space VoxRT/linnet-tts.

1. Check the prerequisites. You need a current Python 3 from the Microsoft Store or python.org. During installation, tick "Add Python to PATH". You don't need a graphics card; the model runs on the processor (CPU). Then open PowerShell, the blue Windows command window.
2. Create a working folder. Create a dedicated folder for the project and switch into it. Run all further commands in this folder.

   mkdir linnet
   cd linnet

3. Create a virtual environment. A virtual environment (venv) is a separate space just for this project, so nothing gets mixed up with other programs. Once activated, the line starts with "(.venv)". If you get an error about the "execution policy", first run Set-ExecutionPolicy -Scope CurrentUser RemoteSigned and confirm with Y.

   python -m venv .venv
   .venv\Scripts\Activate.ps1

4. Install the software. This installs the speech engine linnet-tts, soundfile for saving the WAV file and huggingface_hub for downloading the model files. linnet-tts doesn't need PyTorch, so it's small. When pip reports "Successfully installed", you're good.

   pip install linnet-tts soundfile huggingface_hub

5. Download the model files. This downloads the weights (the model's "knowledge", about 83 MB), the tokenizer (which splits text into building blocks) and the built-in female voice into the local subfolder. You don't need the large fp32 file for this guide. Afterwards the files are in local and local\voices.

   hf download VoxRT/linnet-en-35m linnet-en-35m.fp16.safetensors tokenizer.model voices/linnet-en-f1.npy --local-dir local

6. Create the script. Open Notepad, paste the code below and save it in the linnet folder as demo.py. Make sure Notepad doesn't turn it into demo.py.txt: choose "All files" as the file type when saving.

   import linnet_tts, soundfile as sf
   
   model = linnet_tts.Model("local/linnet-en-35m.fp16.safetensors", "local/tokenizer.model")
   voice = model.load_voice("local/voices/linnet-en-f1.npy")
   audio = model.synthesize("Hello from Linnet.", voice, temp=0.3, cfg=2.0, seed=0)
   sf.write("hello.wav", audio, model.sample_rate)

7. Generate your first speech. Run the script. It only takes a few seconds. Afterwards the folder contains hello.wav, which you can play with a double-click.

   python demo.py

8. Result. Your folder now contains hello.wav, in which a female voice says "Hello from Linnet."

macOS

macOS: not officially tested. There's a speed measurement for macOS on an Apple M5 (Rust runtime), but a ready-made Python package for the Mac isn't explicitly mentioned; according to the source, the macOS part of the iOS package isn't considered a supported product. If pip doesn't work, use the online demo in the Hugging Face Space VoxRT/linnet-tts.

1. Check the prerequisites. You need a current Python 3. The easiest way to check: open Terminal (Applications > Utilities) and type python3 --version. If Python is missing, macOS will offer to install the Xcode command line tools, which include Python. Accept. You don't need a graphics card.
2. Create a working folder. Create a dedicated folder for the project and switch into it. Run all further commands there.

   mkdir linnet
   cd linnet

3. Create a virtual environment. A virtual environment (venv) is a separate space just for this project. Once activated, the line starts with "(.venv)". You need to repeat this activation command in every new Terminal window.

   python3 -m venv .venv
   source .venv/bin/activate

4. Install the software. This installs the speech engine linnet-tts, soundfile for saving the WAV file and huggingface_hub for the download. When "Successfully installed" appears, you're good.

   pip install linnet-tts soundfile huggingface_hub

5. Download the model files. This downloads the weights (about 83 MB), the tokenizer (which splits text into building blocks) and the built-in female voice into the local subfolder. You don't need the large fp32 file here.

   hf download VoxRT/linnet-en-35m linnet-en-35m.fp16.safetensors tokenizer.model voices/linnet-en-f1.npy --local-dir local

6. Create the script. This command writes the small Python program demo.py into your folder. Copy the whole thing into Terminal and press Enter.

   cat > demo.py <<'EOF'
   import linnet_tts, soundfile as sf
   
   model = linnet_tts.Model("local/linnet-en-35m.fp16.safetensors", "local/tokenizer.model")
   voice = model.load_voice("local/voices/linnet-en-f1.npy")
   audio = model.synthesize("Hello from Linnet.", voice, temp=0.3, cfg=2.0, seed=0)
   sf.write("hello.wav", audio, model.sample_rate)
   EOF

7. Generate your first speech. Run the script. After a few seconds, the folder contains hello.wav. Use open hello.wav to play it right away.

   python3 demo.py
   open hello.wav

8. Result. Your folder now contains hello.wav, in which a female voice says "Hello from Linnet."

Linux

Linux: supported. Linux is officially supported (Python package, C library for x86_64 and aarch64, even Raspberry Pi). Everything runs on the processor; no graphics card is used.

1. Check the prerequisites. You need Python 3 with pip and the venv module. Check in the terminal with python3 --version. On Ubuntu/Debian, venv is sometimes missing; install it through the package manager (package python3-venv). You don't need a graphics card or CUDA.
2. Create a working folder. Create a dedicated folder for the project and switch into it.

   mkdir linnet
   cd linnet

3. Create a virtual environment. A virtual environment (venv) is a separate space just for this project. Once activated, the line starts with "(.venv)". You need to reactivate it in every new terminal.

   python3 -m venv .venv
   source .venv/bin/activate

4. Install the software. linnet-tts is the fast Rust runtime packaged for Python, with no PyTorch at all. soundfile saves the WAV file, and huggingface_hub provides the hf download command. "Successfully installed" means it worked.

   pip install linnet-tts soundfile huggingface_hub

5. Download the model files. This downloads the weights (the model's "knowledge", about 83 MB), the tokenizer (which splits text into building blocks) and the built-in female voice into the local subfolder. The large fp32 file is only meant for the PyTorch reference and isn't needed here.

   hf download VoxRT/linnet-en-35m linnet-en-35m.fp16.safetensors tokenizer.model voices/linnet-en-f1.npy --local-dir local

6. Create the script. This command writes the example program from the guide into your folder as demo.py. Copy the whole thing into the terminal.

   cat > demo.py <<'EOF'
   import linnet_tts, soundfile as sf
   
   model = linnet_tts.Model("local/linnet-en-35m.fp16.safetensors", "local/tokenizer.model")
   voice = model.load_voice("local/voices/linnet-en-f1.npy")
   audio = model.synthesize("Hello from Linnet.", voice, temp=0.3, cfg=2.0, seed=0)
   sf.write("hello.wav", audio, model.sample_rate)
   EOF

7. Generate your first speech. Run the script. After a few seconds, the folder contains hello.wav (24 kHz, mono). Open it with any audio player.

   python3 demo.py

8. Result. Your folder now contains hello.wav, in which a female voice says "Hello from Linnet."