Skip to content

v0.5.0

Choose a tag to compare

View all tags
@danielr-ceva danielr-ceva released this 26 Mar 13:40
· 6 commits to main since this release
v0.5.0
37e6b49

DPDFNet v0.5.0

New Features

  • New model: dpdfnet8_48khz_hr - Best-quality high-resolution 48 kHz speech enhancement model.
  • Parallel multi-threaded inference - ONNX, TFLite, and dpdfnet enhance-dir CLI now process files concurrently using thread-pool workers. Each thread gets its own independent runtime session to avoid lock contention. Configurable via --workers (defaults to all CPU cores).
  • CLI banner - enhance and enhance-dir commands now display model info (name, sample rate, description) at startup.
  • Checkpoint loader fallback - ONNX export scripts now gracefully handle both raw state_dict and full checkpoint files containing a state_dict key.

Improvements

  • Per-file timing stats (total time, avg frame latency, RTF) printed for both ONNX and TFLite batch inference.
  • Updated ONNX model sizes in README to reflect latest exports.
  • Fixed missing --model flag in package/README.md example for enhance-dir.

Models

Model Sample Rate Params ONNX Size
baseline 16 kHz 2.31M 8.3 MB
dpdfnet2 16 kHz 2.49M 9.7 MB
dpdfnet4 16 kHz 2.84M 11.1 MB
dpdfnet8 16 kHz 3.54M 13.9 MB
dpdfnet2_48khz_hr 48 kHz 2.58M 10.0 MB
dpdfnet8_48khz_hr 48 kHz 3.63M 14.2 MB

Install / Upgrade

pip install --upgrade dpdfnet
Assets 2
Loading