3 days ago
-
Update runtime, tools, eval scripts and results; add assets/v2-lowram
Runtime: graph file names from manifest, depth loaded before backbone, lang option kept. Tools: make_asset_variant, depth int8 quantizer, backbone quant sweep. Benchmarks: end-to-end eval, speaker similarity, listening pairs, 2026-10-29 results. New optional asset set assets/v2-lowram (int8 depth, ~640 MB less peak RAM, ~2.5x slower).
4 days ago
-
Remove __pycache__ and unrelated ASR tests
-
Initial release: optimized CPU runtime and converted weights of maya2-native
