Original title: KittenML/kitten-tts-2-mini
Uploaded by KittenML on 11 13: 82M parameters, 16 languages, 118 voices, 24 kHz output. Provides two ONNX weights: int8 (82.7 MB) and int4 (47.5 MB), both runnable on CPU without PyTorch, plus float32 PyTorch weights. English has 11 character voices with persona descriptions (historian, radio journalist, sports commentator, etc.); voices with the same name can also speak German, Spanish, Italian, Portuguese, and Russian. Other languages each have 2–8 numbered voices. The model, voice packs, and G2P files are all Apache-2.0.
Why it matters The int4 version is under fifty MB, runs on CPU, and has an Apache license allowing commercial use, making it suitable for offline reading and multilingual dubbing
Upvotes 35 · Likes 35 · Downloads 3887
Reasons
- Covers 16 languages: English, German, Spanish, Italian, Portuguese, Russian, Hindi, Danish, Finnish, French, Japanese, Korean, Dutch, Swedish, Swahili, and Turkish
- Install with one line: pip install kittenml; downloads only the necessary files based on the selected backend and weight format
- Voices can be specified by name or description; use the speed parameter to adjust speaking rate
Usage
- pip install kittenml; after KittenTTS("KittenML/kitten-tts-2-mini"), call generate(text, voice="Eleanor") and write output at 24000 Hz using soundfile
- To switch languages: select a voice for that language, or pass only language="de" to use the first voice for that language; adding language= to an English voice allows it to speak German, Spanish, Italian, Portuguese, and Russian
- Use weights="int4" for the int4 version (requires onnxruntime 1.20 or higher); use backend="torch", device="cuda" to run float32 weights on GPU
Risks
- Does not support Chinese; no audio quality or error rate benchmarks are provided, and there is no voice cloning
- Non-English languages mostly have numbered voices like Female 1 and Male 1, without style descriptions. This entry was not run