Kokoro methodology study mirrors

Kokoro-Hexagon 0a03be39Updated 2026-09-26

The supplied comparative report was used only to identify repositories worth examining. Its performance rankings, latency, footprint, and power figures are not accepted as cross-system benchmarks: they lack a common model/input, hardware, warm-up, timing boundary, and measurement protocol.

The following independent trees were cloned to a local upstream mirror at the immutable commits shown. Inspect with git show <commit>:<path> or the shared search index; do not treat their working trees as Kokoro build inputs. No code, weights, generated models, or binaries from these trees were copied into this repository or its production closure.

Mirror	Upstream	Commit	Methodology question
`Kokoro-FastAPI`	`https://github.com/remsky/Kokoro-FastAPI.git`	`b4ef64b1ce60682debda4fe0a066259e284eb1b4`	How are long requests segmented, queued, and delivered to playback?
`kokoro-mlx`	`https://github.com/gabrimatic/kokoro-mlx.git`	`78e0f87eb3f5105a451bb485a46d01a36364358e`	How is the stock graph scheduled against shared memory?
`kokoro.cpp`	`https://github.com/simonfxr/kokoro.cpp.git`	`a9e31430838c8bdc3d4beaaab62753681b9f7839`	How are weights mapped and intermediate tensors reused?
`kokoro-server`	`https://github.com/marty1885/kokoro-server.git`	`952be751c3c4ee325a8049fa931eae037ef55f3b`	Which operator boundaries are measured on a constrained accelerator?
`sherpa-onnx`	`https://github.com/k2-fsa/sherpa-onnx.git`	`040afe360a38e25daaa325ce8889abf93ea02609`	How are embedded TTS requests, cancellation, and audio chunks bounded?
`kokoro-onnx`	`https://github.com/thewh1teagle/kokoro-onnx.git`	`3596b26764286a7de9d90c363e988d50578918e5`	Which tensor shapes and model variants need independent parity checks?
`kokoro-ios`	`https://github.com/mlalma/kokoro-ios.git`	`4d6d1d8ff8cd012014180c9cd4cf0151e7682354`	How is device playback coordinated with synthesis?
`Kokoros`	`https://github.com/lucasjinreal/Kokoros.git`	`29e99ad5a5aa64b97e1e8e963e6d73b0267d796a`	How is first-audio latency separated from sustained throughput?

These are questions to investigate, not verified findings or implementation instructions. The original pinned Kokoro source, checkpoint, and config remain the only model-behavior authority. PowerShell-authored lowering and direct Hexagon emission remain the independent product path.