iliasaz/perth-coreml

Perth-Net Implicit — CoreML

CoreML conversion of Perth-Net Implicit, the inaudible audio watermarker from Resemble AI's Perth, for iOS 18+ / macOS 15+. The encoder and decoder convolution stacks run as CoreML packages (on the Apple Neural Engine by default); the STFT/ISTFT, magnitude masking and resampling run on the host in Accelerate, since CoreML has no complex-number support.

These files are loaded by the Swift package iliasaz/perth-coreml, and used by iliasaz/chatterbox-coreml to watermark every utterance it generates — as upstream Chatterbox does.

Which file you need

file	precision	size	needed for
PerthEncoder.mlpackage	fp16	4.7 MB	embedding a watermark (applyWatermark) — the only file chatterbox-coreml downloads
PerthDecoder.mlpackage	fp16	14.0 MB	detecting a watermark (getWatermark)
PerthEncoder_fp32.mlpackage, PerthDecoder_fp32.mlpackage	fp32	9.5 MB, 28.0 MB	numeric-parity reference tier — not needed at runtime
perth_assets.json, perth_resampler.json	—	46 KB	DSP constants written alongside the conversion; the Swift runtime embeds the same constants and does not read these

If you only embed, fetch only the encoder:

hf download iliasaz/perth-coreml --include "PerthEncoder.mlpackage/*" --local-dir ./perth

Every package declares Float32 inputs and outputs; the fp16 packages differ only in compute precision. Accuracy against stock Python Perth is documented in the perth-coreml README.

Links

• Upstream — resemble-ai/Perth by Resemble AI.
• Swift runtime — iliasaz/perth-coreml.
• Used by — iliasaz/chatterbox-coreml.

License

MIT. These packages are a CoreML format conversion of the Perth-Net Implicit checkpoint that Resemble AI publishes under the MIT licence (resemble-perth); the conversion inherits those terms, and copyright in the weights stays with Resemble AI — no claim of ownership is made over them here. Redistribute with attribution to Resemble AI.