Table of contents

Model weight admission and synthesis handoff

Kokoro-Hexagon 0a03be39Updated 2026-09-27

The PowerShell readers Read-KokoroAcousticWeights.ps1, Read-KokoroDecoderWeights.ps1, and Read-KokoroGeneratorWeights.ps1 admit the pinned kokoro-v1_0.pth into named tensor maps for the existing acoustic, decoder, and generator stages. Each checks the exact checkpoint length and SHA-256 before parsing, then exact tensor names, shapes, FP32 storage, contiguous strides, and finite values. The three maps contain 171, 72, and 303 tensors respectively. The checkpoint is in ignored build/cache/, not the repository history or APK. Test-KokoroWeightCoverage.ps1 accounts for all 548 checkpoint entries: 546 map into the consumed computation; the other two are ALBERT pooler weight and bias. Kokoro's CustomAlbert.forward returns only last_hidden_state (kokoro/modules.py:180-183 at the pinned revision). The pinned Transformers source 8ac2b916b042b1f78b75c9eb941c0f5d2cdd8e10, src/transformers/models/albert/modeling_albert.py:720-729, forms pooler_output from sequence_output after the encoder and does not feed it back into last_hidden_state. Thus those two tensors do not enter the stock speech output path; they remain in the source checkpoint for identity.

Source identity: Kokoro dfb907a02bba8152ca444717ca5d78747ccb4bec, kokoro/model.py and kokoro/istftnet.py; stock checkpoint revision f3ff3571791e39611d31c381e3a41a3af07b4987; model digest and config in lib/manifest.json and lib/kokoro-v1_0.config.json.

Executable gates passed:

pwsh -NoProfile -File tools/Test-KokoroAcousticWeights.ps1 -CheckpointPath build/cache/kokoro-v1_0.pth
pwsh -NoProfile -File tools/Test-KokoroDecoderWeights.ps1 -CheckpointPath build/cache/kokoro-v1_0.pth
pwsh -NoProfile -File tools/Test-KokoroLearnedGenerator.ps1 -CheckpointPath build/cache/kokoro-v1_0.pth
pwsh -NoProfile -File tools/Test-KokoroWeightCoverage.ps1 -CheckpointPath build/cache/kokoro-v1_0.pth
pwsh -NoProfile -File tools/Test-KokoroPhonemeToPcmContract.ps1

Invoke-KokoroPhonemeToPcm.ps1 now connects the full configured acoustic branches, decoder prelude/core, harmonic source, learned generator, and PCM head for admitted token IDs and a voice row. The contract gate validates its AST and rejection of invalid boundary tokens and incomplete weight maps. Read-KokoroVoiceRow.ps1 also verifies the pinned af_heart.pt artifact and selects stock row phoneme_count - 1; its row-shape and bounds gate passes. Test-KokoroPhonemeToPcmFull.ps1 defines the pending complete stock-layer execution gate and writes output PCM only under ignored build/ after a pass. No completed stock-weight forward has been recorded yet. The older reduced-layer acoustic fixture was deliberately stopped during a slow scalar rerun and is not a new pass. No listening, device execution, numerical parity, latency, or speech claim follows from this handoff. The next promotion gate is a bounded same-input stock numerical differential through the complete chain.