The model based on 'DTLN' project and your paper about 'DTLN-aec' can't be trained as good as your pretrained one.

[steven8274]

Hi, Nils, after carefully reading the paper and the code in 'DTLN' project, I modified the model of DTLN to DTLN-aec's.I check the model structure again and again to make sure that it's coinsident with your paper.Then I compose the trainging data as your describe in the paper except for 'Random spectral shaping' which I'm not sure how to implement.The dataset include farend speech and echo from 'synthetic' dataset and 'real' dataset.It also include my synthesized echo data with 'DNS-Challenge 3' speech data.But the model I got performed not so well as the pretraned one.Could you help me to verify the model structure or the trainging dataset composing process?Thanks in advance!

[leizhu1989]

@steven8274 hello! may I ask how to understand the "All signals used as input to the model are subject to a random gainchosen from a uniform distribution ranging from -25 to 0 dB relative to the clipping point" about "relative to the clipping point" in paper. where "the clipping point" conme from. thank you!

[steven8274]

@steven8274 hello! may I ask how to understand the "All signals used as input to the model are subject to a random gainchosen from a uniform distribution ranging from -25 to 0 dB relative to the clipping point" about "relative to the clipping point" in paper. where "the clipping point" conme from. thank you!

In my opinion, the clipping point means the original signal, which is relative to the adjusted signal processed by the chosen gain.

[leizhu1989]

@steven8274
thank you for you reply , I still don't quite understand what signal to add，example：

import numpy as np
min_gain_dB = -25
max_gain_dB = 0
random_gain_dB = np.random.uniform(min_gain_dB, max_gain_dB)
actual_gain_dB = reference_cut_point + random_gain_dB

is “reference_cut_point” like this: if I get mic signal (Echo signal+near-end signal) value is m，whether the “reference_cut_point” is 20log(m/32767) ,16k sampling rate

[steven8274]

@steven8274 thank you for you reply , I still don't quite understand what signal to add，example：

import numpy as np min_gain_dB = -25 max_gain_dB = 0 random_gain_dB = np.random.uniform(min_gain_dB, max_gain_dB) actual_gain_dB = reference_cut_point + random_gain_dB

is “reference_cut_point” like this: if I get mic signal (Echo signal+near-end signal) value is m，whether the “reference_cut_point” is 20log(m/32767) ,16k sampling rate

Yes, that's just my understanding.Sorry for my late relpy...

[leizhu1989]

@steven8274 thank you for your reply.