simonfxr/kokoro.cpp-GGUF

Kokoro-82M for kokoro.cpp — F16, Q8_0, and all 54 voices

Reproducible GGUF conversions of the official Hexgrad Kokoro-82M v1.0 checkpoint for kokoro.cpp, a standalone C/C++ runtime powered by GGML.

What is included: the reference F16 model, a validated Q8_0 model, and every one of the 54 official v1.0 voice packs across all supported languages. No unofficial or third-party voices are mixed into this repository.

Files

File	Precision	Size	Notes
kokoro-82m-f16.gguf	F16/F32	156.14 MiB	Reference-quality conversion
kokoro-82m-q8_0.gguf	Q8_0/F16/F32	134.78 MiB	111 matrix tensors quantized; sensitive/unsupported tensors retained
voices/kokoro-voice-*.gguf	F32	26.91 MiB total	All 54 official voices; about 510 KiB each
manifest.json	—	—	SHA256, size, source revision, and checkpoint provenance
VALIDATION.md	—	—	Tensor and fixed-seed audio comparison results

Voice packs remain F32 because they are small and are indexed style embeddings, not matrix-multiplication weights. A model file and one voice file are both required for synthesis.

Quick start with kokoro.cpp

# Download one model and one voice.
hf download simonfxr/kokoro.cpp-GGUF \
  kokoro-82m-q8_0.gguf \
  voices/kokoro-voice-af_heart.gguf \
  --local-dir models

# Build the runtime. See the GitHub README for platform/backend options.
git clone --recursive https://github.com/simonfxr/kokoro.cpp.git
cd kokoro.cpp
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

# Synthesize 24 kHz mono audio.
./build/kokoro-cli \
  -m ../models/kokoro-82m-q8_0.gguf \
  -v ../models/voices/kokoro-voice-af_heart.gguf \
  -l en-us \
  -o output.wav \
  "Hello from Kokoro dot C P P."

GPU execution is available when kokoro.cpp is built with a supported Vulkan, Metal, or CUDA backend. GGUF is the container format; llama.cpp and whisper.cpp do not implement this Kokoro architecture.

Languages and voices

Kokoro reports eight languages. Its pipeline distinguishes American and British English, resulting in nine language variants. kokoro.cpp takes the espeak-ng voice shown in the third column, not Kokoro's one-letter pipeline code.

Language	Kokoro code	kokoro.cpp -l	Voices
American English	a	en-us	20
British English	b	en-gb	8
Spanish	e	es	3
French	f	fr-fr	1
Hindi	h	hi	4
Italian	i	it	2
Japanese	j	ja	5
Brazilian Portuguese	p	pt-br	3
Mandarin Chinese	z	cmn	8

See voices/README.md for every voice name and ready-to-copy commands. The machine-readable mapping is in voices/languages.json.

Conversion and provenance

• Base model: hexgrad/Kokoro-82M
• Pinned source revision: f3ff3571791e39611d31c381e3a41a3af07b4987
• Official checkpoint SHA256: 496dba118d1a58f5f3db2efc88dbdc216e0483fc89fe6e47ee1f2c53f18ad1e4
• Runtime and converters: simonfxr/kokoro.cpp
• Output checksums: manifest.json

The conversion pipeline fuses PyTorch weight normalization, preserves Kokoro's sparse 178-entry IPA vocabulary, maps all 459 runtime tensors, and emits GGUF metadata expected by kokoro.cpp. Q8_0 quantizes only eligible two-dimensional matrix weights whose GGML row width supports 32-element blocks. Convolutional kernels remain F16 and rank-one tensors remain F32.

Reproduce these files from the official weights:

git clone --recursive https://github.com/simonfxr/kokoro.cpp.git
cd kokoro.cpp
python3 -m venv .venv-convert
. .venv-convert/bin/activate
pip install -r models/requirements-convert.txt
./models/convert-original.sh

Validation

The generated files were checked for tensor names, shapes, dtypes, metadata, and source hashes. Both models were loaded by kokoro.cpp; F16 and Q8_0 English and Q8_0 Spanish synthesis produced valid 24 kHz WAV files.

Against the previously published cstr files:

• The F16 file has the same size and cosine similarity 0.999999999986 over all 81,731,256 dequantized elements.
• Our Q8_0-versus-F16 relative L2 error is 0.00512913, essentially matching cstr's 0.00512568.
• Shared official voices af_heart, ef_dora, and ff_siwis are exactly equal element-for-element.
• Fixed-seed cstr-F16 versus this F16 output has waveform correlation 0.999920 and log-spectral distance 0.529 dB.

Full methodology, caveats, and pairwise results are in VALIDATION.md. PCM comparisons are phase-sensitive and do not replace multilingual listening tests or perceptual metrics.

Limitations

• Maximum input is 510 raw phoneme tokens because the model adds two boundary tokens to PL-BERT's 512-position context.
• Phonemization is provided by espeak-ng in kokoro.cpp, while the official Python package uses Misaki. Pronunciation can therefore differ.
• Mandarin tone numbers are not represented in Kokoro's 178-token vocabulary.
• espeak-ng Japanese kanji coverage is incomplete; pre-phonemized IPA can be supplied when needed.
• Q8_0 is smaller and validated, but F16 remains the reference choice for maximum fidelity. Lower-bit quantizations are intentionally not published.
• Voice availability does not imply identical quality across languages. Consult the official model card and official voice notes.

Intended use

This repository is for local and embedded text-to-speech inference with kokoro.cpp, including research, accessibility, applications, and deployment permitted by the upstream Apache-2.0 license. Users remain responsible for appropriate disclosure, consent, and compliance when publishing synthetic speech. Do not use generated audio to impersonate people or facilitate fraud.

License and attribution

The Kokoro-82M weights and official voice packs are Apache-2.0 licensed. The GGUF files are numerical conversions of those weights and retain that license. kokoro.cpp has its own software/dependency licensing details; consult its LICENSE and NOTICE before distributing binaries.

Architecture: StyleTTS 2 / ISTFTNet. Original model, training information, datasets, acknowledgements, and citations are documented by Hexgrad.