Stock scalar-reference profile

Scope: one admitted phoneme, boundary token IDs [0,43,0], speed 100, af_heart voice row zero, pinned Kokoro checkpoint SHA-256 496DBA118D1A58F5F3DB2EFC88DBDC216E0483FC89FE6E47EE1F2C53F18AD1E4. These are local Windows PowerShell-reference timings, not Android, cDSP, speaker, or time-to-first-audio measurements.

The acoustic reader's original exhaustive admission took 0.9 s to hash the 312 MiB checkpoint, 1.1 s to parse tensor descriptors, effectively 0 s to check the 171 descriptor shapes/strides, and 113.9 s to copy their bytes and scan every FP32 value for finiteness. The exhaustive acoustic, decoder, and generator reader gates had already passed for this exact checkpoint digest. For development re-runs, -SkipFiniteScan still verifies that exact digest, all declared shapes and strides, and the byte copies; it rejects any change to the audited digest. The acoustic byte-admission stage then took 2.0 s. This is reuse of an audited property of identical bytes, not a claim that unchecked weights are valid.

With the same model input, the full 12-repeat scalar ALBERT encoder took 423.8 s. The following BERT projection took 2.8 s, duration branch 43.1 s, three-layer text encoder 36.2 s, alignment below 0.1 s, and F0/N branch 63.3 s. The decoder prelude completed, then the run was stopped deliberately before the decoder core and generator. No PCM was produced or heard. The PowerShell scalar operator path is therefore an impractical product backend, but its bounded gates remain useful for same-input/same-weight differential checks of directly emitted Hexagon operators.

The source contract for the ALBERT repeat count is pinned Kokoro dfb907a02bba8152ca444717ca5d78747ccb4bec and Transformers 8ac2b916b042b1f78b75c9eb941c0f5d2cdd8e10; the timing applies only to this repository's PowerShell implementation and host. It does not establish a performance bound for the DSP path.