kokoro_parakeet_aio_models
Ready-to-run assets for two standalone C++ CPU inference engines, so neither needs a Python step before it works. Their builds download these files.
- kokoallovero: Kokoro-82M text to speech.
- parakeetogo: parakeet-redux speech to text.
kokoro/ (Apache 2.0)
| file | what |
|---|---|
kokoro.bin, kokoro.tsv |
hexgrad/Kokoro-82M kokoro-v1_0.pth as one flat little-endian f32 blob with weight norm folded, plus a tensor index (name, offset in floats, shape). Same numbers as the original checkpoint. |
voices.bin, voices.tsv |
Voice packs af_heart, am_michael, bf_emma from the same repo, same layout. |
vocab.tsv |
Kokoro's phoneme vocabulary (codepoint, id). |
g2p/lexicon.bin |
English pronunciation lexicon with part-of-speech variants, tokenizer rules and normalization tables. |
g2p/tagger.bin |
Averaged-perceptron part-of-speech tagger that picks homograph readings, distilled from spaCy en_core_web_sm. |
g2p/guesser.bin |
Small character-to-phoneme transformer (f16) for words the lexicon lacks. |
The model blobs are reproducible from the public checkpoint with kokoallovero's tools/export_weights.py. The G2P assets were built from a lexicon and text corpus that are not published; the training code is in kokoallovero's tools/.
parakeet/ (CC-BY-4.0)
model.safetensors and tokenizer.json, unmodified copies from moondream/parakeet-redux, the ternary version of nvidia/parakeet-tdt-0.6b-v3. Mirrored here only so both engines fetch from one place; credit and licence are moondream's and NVIDIA's.