This repository is the new home for the improved DTLN/AEC rewrite formerly hosted at https://github.com/GR33N-WCL/DTLN_improved.git.

It is a ground-up rewrite of the original DTLN codebase and the existing repository-specific AEC branch. The project keeps the core Dual-Signal Transformation LSTM ideas, but turns the repository into a maintainable Python package with explicit configuration, reusable data pipelines, test coverage, conversion helpers, and compatibility shims for the legacy scripts.

The rewrite intentionally keeps the strengths of the original project and the existing AEC branch:

• The real-time two-stage DTLN/AEC inference structure is preserved.
• Existing TFLite model artifacts under DTLNbeta/ remain usable.
• Legacy entry points are still available through compatibility wrappers.
• The repository-specific dual-input AEC workflow is kept instead of being flattened into a denoising-only layout.

What improved around those strengths:

• The code is now package-structured and configuration-driven instead of being centered on one large script.
• Data validation, manifest building, export, inspection, and reporting are now first-class modules.
• Training, export, offline inference, and streaming-style inference share reusable runtime code.
• Regression tests and benchmark scripts are included so changes can be checked instead of guessed.

• The single-file training script was decomposed into a package under src/moneyprint_dtln.
• Both the original single-input denoising model and the repository-specific dual-input acoustic echo cancellation model are supported.
• Dataset indexing, file pairing, frame chunking, and metadata validation now live in dedicated modules.
• Export to SavedModel and TFLite is routed through reusable service classes.
• Offline enhancement and frame-by-frame streaming inference are built on top of shared runtime code.
• The previous DTLNbeta scripts are kept as thin compatibility wrappers so existing workflows do not break immediately.

The repository now includes benchmark scripts and reports for the rewritten runtime path.

• Real TFLite runtime benchmark: the rewritten inference loop is about 6% to 7% faster than the previous loop while producing identical output for the same AEC .tflite weights.
• Quality regression checks: legacy and rewritten loops produce identical audio output on the tested real .tflite models, so the rewrite does not introduce an inference-quality regression by itself.
• The benchmark work also showed that the current checked-in AEC weights are the limiting factor for quality on the tested speech cases; the runtime rewrite improves execution efficiency, not the learned model quality.

src/moneyprint_dtln/
  config.py            typed project configuration
  data/                manifests, file pairing, and tf.data creation
  models/              DTLN denoiser and AEC model builders
  training.py          trainer and callbacks
  exporters.py         SavedModel and TFLite export
  runtime.py           offline and streaming inference
  cli.py               command line entry point
tests/
  lightweight tests for config, manifests, and audio framing helpers
DTLNbeta/
  legacy compatibility wrappers

• Usage guide
• Data conventions
• Testing guide
• Benchmarking notes
• Architecture overview
• Rewrite evidence checklist
• Contribution guide
• Changelog

Install the package in editable mode:

pip install -e .

Inspect the available commands:

python -m moneyprint_dtln.cli --help

Create a dataset manifest:

python -m moneyprint_dtln.cli manifest \
  --config configs/aec_train.json \
  --root /path/to/dataset \
  --output dataset_manifest.json

Train the AEC model:

python -m moneyprint_dtln.cli train \
  --config configs/aec_train.json

Enhance a folder offline:

python -m moneyprint_dtln.cli enhance \
  --config configs/aec_infer_tflite.json \
  --input /path/to/input \
  --output /path/to/output

The denoising mode expects parallel noisy and clean .wav files with the same relative path under separate roots.

The AEC mode expects triplets that can be identified by a shared suffix:

• nearend_mic_fileid_<id>.wav
• farend_speech_fileid_<id>.wav
• nearend_speech_fileid_<id>.wav

The pairing logic is configurable if your naming rules differ.

• TensorFlow is loaded lazily so utility modules and tests can run without a full training environment.
• The repository still contains historical model artifacts under DTLNbeta/. They are not required for the new package layout.