soniqo/LocalVQE-v1.4-AEC-200K-ONNX-FP32

LocalVQE v1.4-AEC 200K ONNX

Portable Float32 ONNX export of LocalVQE v1.4-AEC's stateful neural residual echo mask. This is a hybrid echo canceller, not a standalone waveform model. The exact adaptive front end and spectral codec remain host DSP.

Contract

Property	Value
Sample rate	16 kHz mono
FFT / hop	512 / 256 samples
Frame interval	16 ms
Model size	0.96 MiB
Neural input	residual and echo-estimate spectra, each [1,2,1,256]
Neural output	enhanced spectrum [1,2,1,256]
State	seven explicit Float32 input/output tensor pairs

The host must run the released GCC-PHAT prealignment, partitioned-block frequency-domain Kalman filter, 2,742-weight controller, sqrt-Hann STFT, and overlap-add. Passing microphone and playback spectra directly to the ONNX model does not implement LocalVQE.

Validation

The exporter rejects a graph unless twelve recurrent frames agree with the pinned PyTorch implementation within a relative RMSE of 1e-4.

Measurement	Result
PyTorch/ONNX relative RMSE	1.605e-06
Warm CPU inference per 16 ms frame	0.539 ms

Latency is measured on the export host and is not a Windows hardware claim.

C++ usage

The complete native implementation is provided by speech-core:

#include <speech_core/models/onnx_localvqe_echo_canceller.h>

speech_core::OnnxLocalVQEEchoCanceller aec(
    "/path/to/soniqo/LocalVQE-v1.4-AEC-200K-ONNX-FP32");
aec.process_frame(
    microphone.data(), reference.data(), cleaned.data());

The two input frames must describe the same capture interval. Reset the model whenever either stream is discontinuous.

Files

File	Description
LocalVQEAECResidualMask.onnx	Functional recurrent residual-mask graph
LocalVQEAECFrontend.json	Exact controller weights, sqrt-Hann window, and DSP geometry
config.json	Graph contract, provenance, hashes, and parity
README.md	This model card

Source and limitations

Converted from the Apache-2.0 localvqe-v1.4-aec-200K-f32.gguf release, run v1.4.r005. The source SHA-256 is recorded in config.json.

• This model removes playback echo; it is not a general noise suppressor.
• A synchronized playback reference is mandatory.
• Reset on capture gaps, device changes, or timestamp discontinuities.
• Evaluate the complete host DSP and model together on the intended hardware.

Links

• LocalVQE
• speech-core
• C++ docs