iky1e/DPDFNet4-CoreML

DPDFNet-4 Core ML

Stateful Core ML and MLX assets for DPDFNet-4, CEVA's 16 kHz real-time speech-enhancement model. This repository is the default DPDFNet-4 source for the DeepFilterNetCoreML Swift product.

Origin

• Official implementation: ceva-ip/DPDFNet
• Official models: Ceva-IP/DPDFNet
• Paper: DPDFNet
• Swift runtime and conversions: kylehowells/DeepFilterNet-mlx
• Conversion script: Scripts/Conversion/convert_dpdfnet_to_coreml.py

Runtime contract

Property	Value
Sample rate	16,000 Hz
Input hop	160 samples / 10 ms
FFT	320 samples
DPRNN blocks	4
Fixed algorithmic delay	800 samples / 50 ms

Recommended model

DPDFNet4-Streaming-FP32.mlpackage is the production default. Its conversion validation against the PyTorch graph produced 0.9999999999 output correlation. The FP16 explicit-state and resident-state variants are included for device-specific experiments but are not selected automatically because their fidelity is lower.

Current 60-second Swift/Core ML validation: 13.953 seconds total (4.30x real time), with 0.99999999 waveform correlation to official ONNX. Live p50/p95 model processing latency was 3.686/4.171 ms on the development Apple Silicon Mac. Device results vary.

Swift usage

import DeepFilterNetCoreML

let enhancer = try await DeepFilterNetCoreMLStreamer.load(
    configuration: .init(variant: .dpdfNet4)
)
let output = try enhancer.processHop(input160Samples)

Files

• DPDFNet4-Streaming-FP32.mlpackage: recommended explicit FP32 state graph.
• DPDFNet4-Streaming.mlpackage: explicit FP16 state graph.
• DPDFNet4-Streaming-State.mlpackage: Core ML MLState graph for supported OS versions.
• DPDFNet4-initial-state-f32.bin: canonical recurrent-state initialization.
• config.json and model.safetensors: matching Swift MLX/DSP model assets.
• conversion-report.json: graph-level validation and measured conversion latency.

License

Apache-2.0, matching the official CEVA DPDFNet repository. See LICENSE and the original project.