The decoder swap passed its word-error gate

Measure the proposed component swap

Both routes used the same community 6-bit MLX T3 generator. The hybrid replaced only the audio stage with the unchanged published ONNX conditional decoder. No CoreML execution occurred in this hybrid.

Every comparison checked the generated speech-token hash. Each MLX baseline was paired with two hybrid outputs using fresh internal noise. Thirty-eight of forty pairs tied on word errors; two hybrid outputs had two fewer edits each.

The cost of the swap

For the generic corpus, median warm time changed from 1.111 to 2.069 seconds and median RTF from 0.230 to 0.411. Selected-voice medians changed from 0.954 to 1.799 seconds. Instrumentation placed the extra time in audio decoding, rather than T3 generation.

The generic corpus had 8/250 baseline edits. Repeating each baseline score across its two matched hybrid comparisons gives 16/500 versus 12/500. Those equal denominators avoid implying that twice as many hybrid clips provide twice as much independent coverage.

The result I can use

The optional route passed its local intelligibility gate, while MLX remained faster. Reference voice audio and conditioning are excluded from this publication. Human preference, multilingual speech and longer-form behavior remain open.

Evidence and publication boundary

This is an edited account of the recorded experiment. The linked archive is private; the methods, aggregate findings and limits are included above. Reference recordings, raw conversations, authentication material and workstation configuration are excluded.

• Source record 1