Test builds only, for checking the Neural Engine compile on M1/M2/M3. main is unchanged.
Issue: https://github.com/FluidInference/FluidAudio/issues/1141
The shipped 560/1120/2240 ms multilingual encoders and the encoder-split/2240ms body halves, with slice_by_index(concat(cache, x)) rewritten to slice the concat's inputs. weight.bin is unchanged. Outputs are byte-identical to main on CPU and ANE (M5 Pro). See test-1141/README.md.
Out of <model>/2240ms/ so default variant downloads (which fetch every .mlmodelc in the tier folder, incl. released SDKs) don't pull the split pieces. The encoder is byte-identical across latin and multilingual, so one copy serves both. https://github.com/FluidInference/FluidAudio/issues/1116
Add opt-in split 2240ms encoder (front end + layers 0-11 + 12-23) for older ANEs (FluidAudio #1116)
Same weights/quantization as encoder.mlmodelc (834 weight blobs bit-identical; chained outputs identical over 38 streamed chunks). The single 2240ms encoder fails ANE register allocation on M3 Pro (falls back to CPU: 238 s load, +2 GB). Issue: https://github.com/FluidInference/FluidAudio/issues/1116
Correct license to OpenMDW-1.1 (commercial-ok) and fix base-model provenance
Resolves the license/provenance contradictions in FluidAudio issue #750.
- Frontmatter: license_name/license_link pointed at the NVIDIA Software and Model Evaluation License (non-commercial); corrected to OpenMDW-1.1, matching the base model nvidia/nemotron-3.5-asr-streaming-0.6b (license: other / openmdw-1.1, ready for commercial use).
- Body 'License & attribution': replaced the NVIDIA Eval License reference with OpenMDW-1.1; states commercial use and redistribution are permitted.
- metadata.json (all 8 bundles): 'model' field repointed from the internal NeMo checkpoint name 'nvidia/nemotron-asr-streaming-multilingual-0.6b' to the public repo id 'nvidia/nemotron-3.5-asr-streaming-0.6b'. Informational only; not read by the Swift loader.