kitten-inference

Fast text-to-speech with Kitten TTS 2 on NVIDIA and AMD GPUs and Kitten TTS 0.8 on the CPU. Install with pip; no PyTorch needed.

On smaller GPUs the Kitten TTS 2 lead is smaller: first audio after 22 ms and complete files 4.5 to 6.9 times faster on an RTX 4090, 58 ms and 2.4 to 4.3 times on a 12 GB RTX 3060. Word error rate on short texts is 0.02 to 0.05% for every engine compared.

Install

Both models use from kitten_inference import KittenTTS with an explicit model name (API).

The commands below start a virtual environment and install support for both model families. On Ubuntu and Debian, python3 -m venv needs sudo apt-get update && sudo apt-get install -y python3-venv first. On Windows, run the blocks in a Command Prompt (in PowerShell, type cmd first), with py -m venv .venv and then .venv\Scripts\activate.bat in place of the first line.

Kitten TTS 2: NVIDIA and AMD GPUs

python3 -m venv .venv && . .venv/bin/activate
python -m pip install -U pip
pip install "kitten-inference"
kitten-tts --model KittenML/kitten-tts-2 --text "Hello from Kitten TTS 2." --voice Bruno --output hello.wav

• NVIDIA GPU on Linux x86-64 (glibc 2.28 or newer: Ubuntu 20.04, Debian 10, RHEL 8 or later) or Windows x64, with driver R570 or newer.
• Python 3.9 to 3.14 (Python 3.15 needs a C compiler until PyYAML, a dependency, ships ready-made 3.15 packages). First use downloads the model (about 4.6 GB) and prepares it once, in about 1 to 3 minutes; later starts reuse it.
• AMD Radeon RX 7000 and RX 9000: supported inside WSL2 on Windows 11, experimental on Linux x86-64 without WSL2. Install ROCm 7 first (WSL2: step 1 of AMD Radeon; Linux without WSL2: AMD's ROCm guide for your distribution), then run the same commands: the package finds the Radeon (on a PC that also has an NVIDIA GPU, add --backend rocm). Real-time factor about 0.06 on an RX 9070 XT (10 s of speech in about 0.6 s). No voice cloning on Radeon.
• Voice cloning (NVIDIA only) needs FFmpeg (sudo apt-get update && sudo apt-get install -y ffmpeg; on Windows winget install --id Gyan.FFmpeg -e, then open a new Command Prompt and activate the venv again; Windows Server 2022 needs winget installed first) and about 11,300 MiB (11 GB) of free GPU memory while the voice is built: on a 12 GB card, other programs may use at most about 500 MiB.
• GeForce GTX 10, GTX 16 and RTX 20 series: supported, not optimal: they have no bf16 tensor cores, so part of the math runs in software (real-time factor 0.22 to 0.37 for complete files on desktop cards with 6 GB or more, against 0.05 to 0.08 on RTX 30, 40 and 50 series cards).
• Without Python: the engine is also a C library and the kitten-tts-native command. Build it with git clone --branch v0.2.0 https://github.com/KittenML/kitten-inference.git, then cd kitten-inference/engines/tts2, and follow engines/tts2/README.md.

The package picks its memory settings from the free GPU memory (GPU memory):

Kitten TTS 0.8: the CPU

python3 -m venv .venv && . .venv/bin/activate
python -m pip install -U pip
pip install "kitten-inference"
python -c "from kitten_inference import KittenTTS; KittenTTS('KittenML/kitten-tts-nano-0.8').generate_to_file('Hello from Kitten TTS.', 'hello.wav', voice='Jasper')"

• Supported: Python 3.8 to 3.14 on Linux x86-64 and ARM64, Windows x64 and Apple Silicon Macs (Python 3.15 needs a C compiler until PyYAML ships ready-made 3.15 packages). Experimental: Intel Macs (the macOS build uses one CPU core), Windows ARM64 (no ARM vector kernels, about 9 times slower than Linux ARM64) and Android ARM64 (platforms).
• On Windows, if KittenTTS() reports "espeak not installed", install the Microsoft Visual C++ Redistributable: winget install Microsoft.VCRedist.2015+.x64.
• On Windows, keep the virtual environment's path under about 150 characters; longer paths stop TTS 0.8 with an error that says so.
• Supported, not optimal:
  • int8 models on Intel CPUs: the int8 kernels for x86 are not tuned yet, so they take 1.2 to 2.2 times the ONNX Runtime time;
  • int8 models on Apple Silicon: the macOS build uses one CPU core, so they take 1.4 to 1.8 times the ONNX Runtime time;
  • creating a model again on the Mac GPU (Metal): each new KittenTTS keeps about 290 MB (nano) to 470 MB (mini) until the process ends, because the engine does not free its GPU copies of the weights yet. Create each model once and reuse it.

Use from Python

from kitten_inference import KittenTTS

tts = KittenTTS("KittenML/kitten-tts-2")
tts.save_wav("hello.wav", tts.generate("Hello from Kitten TTS 2.", voice="Bruno"))
for chunk in tts.stream("Streaming speech, chunk by chunk.", voice="Bruno"):
    ...

from kitten_inference import KittenTTS

KittenTTS("KittenML/kitten-tts-nano-0.8").generate_to_file("Hello from Kitten TTS.", "hello.wav", voice="Jasper")

Docs

• Python API and kitten-tts command: voices, streaming, voice cloning, saved voices, and model options.
• Kitten TTS 2 engine: speed tables, supported GPUs, building from source, the command-line tool and the C API. Also: Windows, AMD Radeon, GPU memory, speed and audio by GPU group.
• Kitten TTS 0.8: models and voices, options, platforms, and upgrading from 0.1.x (import model_inference needs kitten-inference==0.1.1).

License

KittenML's code is Apache-2.0 (LICENSE); the package also contains third-party code under other licences (NOTICE, engines/tts2/NOTICE, engines/tts08/NOTICE). The models download on first use and keep their publishers' licences. The Kitten TTS 2 model is under the Stellon Labs Community License, not Apache or MIT: it limits commercial use, training on the outputs and redistribution, so read it before you build on Kitten TTS 2. The Kitten TTS 0.8 model licences are in engines/tts08/README.md.