Denoiser
Denoiser component that implements the TinyDenoiser as an Audio Framework component. For more information on the TinyDenoiser see https://github.com/GreenWaves-Technologies/tiny_denoiser/
The component processes single channel floating point samples at 16kHz, applies the RNN filter and reconstruct the cleaned signal via overlap-and-add. The noisy signal is windowed (frame size of 25 msec with an hop lenght of 6.25 msec and Hanning windowing) and the STFT is computed. The RNN is fed with the magnitude of the STFT components and return a suppression mask. After weighting, the inverse STFT returns a cleaned audio clip.
Model and STFT settings
The model and STFT has been build with AutoTiler using these settings
H_STATE_LEN = 256 FRAME_SIZE = 400 FRAME_STEP = 100 FRAME_NFFT = 512
QUANT_TYPE = “mixedne16fp16” NN_TYPE = f16 RNN_TYPE_H = uint8_t RNN_TYPE_C = int16_t ZERO = 128
Inputs/Outputs
input- Single channel floating point input buffer containing the signal to be denoised.Type: audio_buffer, float
output- Single channel floating point output buffer containing the denoised signal.Type: audio_buffer, float
IMPORTANT LIMITATIONS
PC model is not included - this will be added in a later version
Tests
Both nonreg.py and gaptest are supported and can be run as usual.
Parameters
enable- Enable/disable parameter for the denoiser. If disabled the input will be copied to the outputType: IntegerRange: [0, 1]