Hugging Face's logo

iky1e
/
DPDFNet8-48kHz-HR-CoreML

Core ML
Safetensors
audio
speech-enhancement
noise-suppression
streaming
dpdfnet
48khz
Model card Files
xet
 Community
DPDFNet8-48kHz-HR-CoreML / README.md
iky1e's picture
iky1e
Publish validated DPDFNet-8 48 kHz HR Core ML models
c92bc18 verified
preview code
|
Raw History Blame Contribute Delete
2.64 kB
metadata 
license: apache-2.0
library_name: coreml
tags:
  - coreml
  - audio
  - speech-enhancement
  - noise-suppression
  - streaming
  - dpdfnet
  - 48khz

DPDFNet-8 48 kHz HR Core ML

Stateful Core ML and MLX assets for CEVA's high-resolution DPDFNet-8 48 kHz HR speech-enhancement model. This is a separate architecture/profile from DPDFNet-4 and is supported by the DeepFilterNetCoreML Swift product.

Origin

Runtime contract

Property Value
Sample rate 48,000 Hz
Input hop 480 samples / 10 ms
FFT 960 samples
DPRNN blocks 8
Fixed algorithmic delay 2,400 samples / 50 ms

Recommended model

DPDFNet8-48kHz-HR-Streaming-FP32.mlpackage is the production default. Conversion validation against PyTorch produced effectively perfect output correlation (0.999999999998). FP16 explicit-state and resident-state variants are included for experimentation but are not selected automatically because their fidelity is lower.

Current 60-second Swift/Core ML validation: 38.664 seconds total (1.55x real time), with 0.99999966 waveform correlation to official ONNX. Live p50/p95 model processing latency was 6.326/7.067 ms on the development Apple Silicon Mac. Device results vary.

Swift usage 

import DeepFilterNetCoreML

let enhancer = try await DeepFilterNetCoreMLStreamer.load(
    configuration: .init(variant: .dpdfNet8HighResolution)
)
let output = try enhancer.processHop(input480Samples)

Files

  • DPDFNet8-48kHz-HR-Streaming-FP32.mlpackage: recommended explicit FP32 state graph.
  • DPDFNet8-48kHz-HR-Streaming.mlpackage: explicit FP16 state graph.
  • DPDFNet8-48kHz-HR-Streaming-State.mlpackage: Core ML MLState graph.
  • DPDFNet8-48kHz-HR-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.