dignome/kitten_tts2

Kitten TTS 2 for audio.cpp

Powered by Stellon Labs.

Self-contained multilingual GGUF exports of KittenML/kitten-tts-2 weights at revision f2eee74f4467855ad58fe16f1fb7b7df473263fd. Each file includes an S3 MeanFlow decoder, speaker encoders, tokenizer, and all 48 prepared voice entries.

Use an audio.cpp build containing the updated Kitten TTS 2 BF16/Q8/Q4 support. The model family is kitten_tts2. BF16/Q8/Q4 model-manager installations share models/kitten-tts2-v2/, with a distinct filename for each precision. Leave kitten_tts2.weight_type unset, or use native, to preserve each file's stored precision. Inference needs no Python or LibTorch installation.

Downloads

Q8 is the default and recommended package. Q4 is experimental.

File	LM body / embedding	Size
kitten-tts2-bf16-multilingual.gguf	BF16 / BF16	4,603,407,040 bytes
kitten-tts2-q8-v2-multilingual.gguf	Q8_0 / BF16	3,282,201,248 bytes
kitten-tts2-q4-multilingual.gguf	Q4_0 / Q8_0	2,273,735,456 bytes
kitten-tts2-nvfp4-multilingual.gguf	NVFP4 / Q8_0	2,273,736,960 bytes
kitten-tts2-tq2_1-default-multilingual.gguf	TQ2_1 / Q8_0, default decoder	1,855,355,872 bytes
kitten-tts2-tq2_1-student-w8-multilingual.gguf	TQ2_1 / Q8_0, student W8 decoder	1,437,030,016 bytes
kitten-tts2-tq2_1-student-w4-multilingual.gguf	TQ2_1 / Q8_0, student W4 decoder	1,413,363,424 bytes

Verify downloads with SHA256SUMS. Names describe LM storage, not the precision of every tensor in the file.

NVFP4 and TQ2_1 require the new package-support branch. The WebUI labels are GGUF NVFP4, GGUF TQ2_1 (CPU) for the default decoder, GGUF TQ2_1 (CPU w4) for the W4 student, and GGUF TQ2_1 (CPU w8) for the W8 student. Their model-manager folders are models/kitten-tts2-nvfp4/, models/kitten-tts2-tq2_1-default/, models/kitten-tts2-tq2_1/, and models/kitten-tts2-tq2_1-w8/, respectively. W8 requires a build with W8 support; earlier versions of the package-support branch do not recognize it.

NVFP4 uses the existing GGML NVFP4 CUDA kernels and the default decoder. Its body-weight reconstruction error is approximately 2.3954% relative L2, not a speech-quality score. All body projections retain NVFP4 on CUDA. Local testing uses an RTX 4060 Ti; other generations have not been validated here.

TQ2_1 uses this branch's type ID 44 and CPU kernels; it is not compatible with stock audio.cpp or Kitten's separate type-43 GGUF files. Select --backend cpu. Its ternary body storage is exact relative to the supplied BF16 checkpoint. The default-decoder package retains the original S3 MeanFlow decoder and its voice conditioning. The W4 student package stores 474 decoder matrices in packed W4 and runs 426 linears as INT8 on CPU automatically; 48 convolutions use F32. Both preset speech and cloning use the decoder selected by the package. These packages do not use the upstream emb4 embedding.

The W8 package uses the shipped frozen one-step student_w8 decoder. Its 457 remaining quantized matrices are stored as INT8 codes with F32 per-channel scales, reconstructing the exported weights exactly. The frozen time-conditioning constants are retained. At runtime, 409 linears use INT8; 48 convolutions use F32. It supports preset voices and cloning, retains all 48 voice entries, and needs no Python or LibTorch at runtime. These checks do not establish better speed or perceptual quality than W4.

Q8 and Q4 preserve the supplied ternary body weights exactly. Q4 quantizes the embedding directly from full BF16 to Q8_0, with about 0.44% relative L2 weight error. This is a weight reconstruction measurement, not an audio quality score. On CUDA, Q4 down projections load as BF16 to avoid non-finite computation, adding about 483 MiB compared with retaining those projections as Q4.

Speech and cloning

audiocpp_cli --family kitten_tts2 --model models/kitten-tts2-v2/kitten-tts2-q8-v2-multilingual.gguf --backend cpu --threads 8 --task tts --voice-id Bruno --text "Hello from Kitten TTS 2." --out hello.wav

Use --backend cuda for a CUDA-enabled build and a package other than TQ2_1. For custom voice cloning:

audiocpp_cli --family kitten_tts2 --model models/kitten-tts2-v2/kitten-tts2-q8-v2-multilingual.gguf --backend cpu --threads 8 --task clon --voice-ref reference.wav --reference-text "The exact words in the reference recording." --text "The new words to say." --out cloned.wav

Cloning requires a 1–30 second reference recording and its transcript. Preset synthesis needs no separate voice recordings. Output is mono 24 kHz.

The package includes 47 upstream presets and the PreparedBruno alias. Named English voices include Bruno, Bella, Jasper and Luna. Language presets cover Arabic, Chinese, French, German, Hindi, Italian, Portuguese, Russian and Spanish. Choose a matching preset or reference clip for the target language. Generation checks cover these ten languages including English; they do not establish native-speaker pronunciation quality.

Validation and limitations

CPU and NVIDIA CUDA synthesis and cloning are checked locally. Q4 remains experimental: longer samples sometimes omitted words. Broader listening review is needed before relying on it for long-form speech. File sizes do not describe runtime RAM or VRAM requirements.

TQ2_1 offers default, student_w4 and student_w8 decoders; the other exports use the default decoder. The upstream TL2/emb4 embedding format is not included. Long chunked speech can omit words with the default or student decoder. Streaming and automatic reference transcription are not supported. Write numbers and abbreviations as spoken words when needed.

Sources and license

• Kitten weights: KittenML/kitten-tts-2, revision f2eee74f4467855ad58fe16f1fb7b7df473263fd.
• Default S3 decoder: ResembleAI/chatterbox-turbo, revision 749d1c1a46eb10492095d68fbcf55691ccf137cd.
• Speaker encoder: the Kitten-distributed speaker model, attributed upstream to pyannote/embedding.

Model materials retain the Stellon Labs Community License in LICENSE. Bundled components retain their respective terms in native/LICENSE and speaker/LICENSE. See NOTICE for attribution and conversion details.

Conversion and native integration by dignome. Prepared with AI assistance and validated locally.