Kokoro-82M — GGUF (ggml-quantised)

GGUF / ggml conversion of hexgrad/Kokoro-82M for use with CrispStrobe/CrispASR.

Kokoro-82M is an open-weight 82 M-parameter TTS model built on the StyleTTS2 architecture. It is multilingual at the model level (178-symbol IPA vocab covering en, es, fr, hi, it, ja, pt, zh) and ships per-language voice packs in hexgrad/Kokoro-82M/voices. Apache-2.0 weights.

Files

File	Quant	Size	Notes
`kokoro-82m-f16.gguf`	F16	156 MB	Reference quality — bit-true to the PyTorch model on the deterministic stages
`kokoro-82m-q8_0.gguf`	Q8_0	135 MB	Recommended — ASR roundtrip identical to F16

Q4_K is not published: per the crispasr-diff kokoro harness it falls below quality bar (cosine similarity collapses to ~0.1 on audio_out and the German backbone produces unintelligible output). Q8_0 saves ~13 % on disk over F16 with no observable quality loss.

Quick start

# 1. Build CrispASR
git clone https://github.com/CrispStrobe/CrispASR
cd CrispASR
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j --target crispasr

# 2. Pull the model + a voice pack
huggingface-cli download cstr/kokoro-82m-GGUF kokoro-82m-q8_0.gguf --local-dir .
huggingface-cli download cstr/kokoro-voices-GGUF kokoro-voice-af_heart.gguf --local-dir .

# 3. Synthesise
./build/bin/crispasr --backend kokoro \
    -m kokoro-82m-q8_0.gguf \
    --voice kokoro-voice-af_heart.gguf \
    --tts "Hello, this is a test of the English phonemizer." \
    --tts-output hello.wav

24 kHz mono WAV. --language <code> switches the in-process libespeak-ng phonemizer (en, es, fr, it, pt, hi, ja, zh — and de/ru when paired with the German backbone).

Voice packs

Voices ship in the sister repo cstr/kokoro-voices-GGUF so the model and per-speaker style files can be versioned independently. CrispASR auto-routes by language when both files sit in the same directory.

Quality verification

ASR roundtrip via parakeet-tdt-0.6b-v3 -l en, voice af_heart:

Quant	Synthesised text	Parakeet output
F16	"Hello, this is a test of the English phonemizer."	"Hello this is a test of the English Phone Miza."
Q8_0	(same text)	"Hello this is a test of the English Phone Miza."

Both quants produce ASR-identical output. The trailing "Phone Miza" is parakeet mis-transcribing the rare word "phonemizer", not a TTS artefact.

crispasr-diff kokoro against the PyTorch reference:

Stage (selection)	F16 cos_min	Q8_0 cos_min
token_ids → durations	1.000	1.000
`pred_lstm_out`	0.999	0.997
`dec_decode_3_out`	0.9999	0.9990
`audio_out` (RNG-divergent)	0.99	0.85

F16 hits 16/16 PASS at cos≥0.999. Q8_0 hits 9/16 PASS at the same threshold; the failed stages stay above cos≥0.85 and don't perturb ASR roundtrip.

Architecture

Component	Details
Text encoder	178-IPA-symbol embedding → 1× CNN block → bidirectional LSTM
pl-BERT	12-layer transformer, 768 d, 12 heads (text prosody features)
Predictor	Duration LSTM + F0/N curves
Decoder (iSTFTNet)	3-stage upsample (10×6 = 60×) + ResBlocks + iSTFT (n_fft=20, hop=5)
Style dim	256 (split as 128+128 between predictor and decoder)
Audio	24 kHz mono output; predictor consumes phoneme IDs

Conversion

python models/convert-kokoro-to-gguf.py \
    --input hexgrad/Kokoro-82M \
    --output kokoro-82m-f16.gguf \
    --outtype f16

build/bin/crispasr-quantize kokoro-82m-f16.gguf kokoro-82m-q8_0.gguf q8_0

The converter handles both legacy weight_g/weight_v and the modern parametrizations.weight.original{0,1} WeightNorm forms, so it also runs on community re-trains like the dida-80b German backbone.

Attribution

Original model: hexgrad/Kokoro-82M (Apache-2.0). hexgrad et al.
Architecture: StyleTTS2-LJSpeech by Yinghao Aaron Li.
GGUF conversion + ggml runtime: CrispStrobe/CrispASR — see src/kokoro.cpp and models/convert-kokoro-to-gguf.py.

License

Apache-2.0, inherited from the base model.

Provenance and EU AI Act Art. 53 note

Upstream model: hexgrad/Kokoro-82M — published by hexgrad.
Upstream licence: apache-2.0. This repository redistributes under the same terms; it grants no rights the upstream licence does not.
What was done here: format conversion and/or quantisation only (GGUF/GGML). No training, no fine-tuning, no merging, no distillation, no change to architecture, vocabulary or capability. Only the numeric representation of the upstream weights differs.
Training data: documented — where it is documented at all — by the upstream provider; see the upstream model card. No training data was used, added or selected by this repository.
Provider status: under Regulation (EU) 2024/1689 the upstream authors remain the provider of this model. Converting the serialisation format does not make this repository the provider of a new general-purpose AI model, and no such claim is made. Questions about training content, copyright policy or model capability belong upstream.