This repository publishes the frozen Kokoro 82M English runtime bundle used by
Whisperstream. It combines a pinned FP32 Kokoro
ONNX graph, nine curated American English voice presets, English lexicons, an
exported neural grapheme-to-phoneme fallback, and English tagger data.
The archive is byte-identical to Whisperstream's R2 fallback object.
tar -xjf .\whisperstream-kokoro-82m-fp32-en-us-2026-07-28.tar.bz2
The installed payload is 345,395,578 bytes.
eSpeak-free scope
The distributed archive contains no eSpeak or espeak-ng executable, library,
Python package, or model. Whisperstream's production inference path does not
invoke eSpeak, espeak-ng, phonemizer, piper-phonemize, num2words, Torch,
Transformers, spaCy, NLTK, Misaki, or kokoro-onnx.
This statement describes this curated runtime bundle. It does not make a claim
about every tool or dependency used by every upstream Kokoro implementation.
Contents
The archive includes:
Kokoro v1.0 FP32 ONNX inference graph
Nine curated American English voice presets
Misaki US gold and silver pronunciation lexicons
Reproducibly exported English neural G2P ONNX fallback
English averaged perceptron tagger data
Exact component license notices and conservative training-data attribution
Export environment, parity report, and per-file source/hash manifest
The archive's export-manifest.json records exact source revisions, sizes,
hashes, transformations, runtime files, and forbidden runtime dependencies.
See NOTICE.txt for the component-by-component provenance and
modification record.
Licenses and responsible use
This is a mixed-license redistribution bundle, not a relicensing of Kokoro.
The Kokoro model and voice materials, Misaki lexicons, and PeterReid G2P source
are under Apache License 2.0. ONNX conversion lineage and NLTK tagger materials
carry MIT notices. Exact texts are in LICENSE and licenses/.
Kokoro's model card identifies additional training-data sources. Conservative
attribution is retained in
licenses/TRAINING-DATA-ATTRIBUTION.txt.
The source datasets themselves are not redistributed here.
Users remain responsible for evaluating fitness, applicable law, speaker and
personality rights, and deployment risks for their own use. No endorsement by
upstream authors, dataset creators, speakers, publishers, or licensors is
stated or implied.
