Skip to content

armgit / Cap

Public Imported from GitHub
0

Open source Loom alternative. Beautiful, shareable screen recordings.

1.2 KB
Raw Blame

Cap RNNoise

This crate embeds the RNNoise model and inference code used by Studio Sound. The C inference code is optimized in development and release builds. Model weights are immutable shared storage; each microphone channel owns its processing state.

Upstream: https://github.com/xiph/rnnoise, revision 70f1d256acd4b34a572f999a05c87bf00b67730d.

The upstream model archive is identified by SHA-256 0a8755f8e2d834eff6a54714ecc7d75f9932e845df35f8b59bc52a7cfe6e8b37. vendor/weights.bin is generated by upstream's dump_weights_blob executable and has SHA-256 1ad07b428be34c74d9678d05b3c387b73ccf0e6203a1a5f9ce58689b30b303c8. The blob uses upstream's little-endian format, matching the supported desktop targets.

The headers and inference sources in vendor are copied from that revision. vendor/src/rnnoise_data.c is generated with unifdef -DUSE_WEIGHTS_FILE src/rnnoise_data.c, retaining the model initializer without committing the 78 MB text representation of the weights. Do not edit the generated initializer or model blob by hand. Refresh them together from the same upstream model.

The upstream BSD license and copyright notices are retained in vendor/COPYING and the source files.