train a model using 24kHz about lpcnet

[jmvalin]

The correct way to do it involves adding a few bands and changing a few constants throughout the code. It's not fundamentally hard, but easy to forget a few things. That being said, considering that 24 isn't that big a change compared to 16, you may be able to get away with pretending that your 24 kHz audio is really 16 kHz and not modifying the code at all.

[luvpine]

Thank you for your answer.
Then, did you normalize training wav files (like -26db, -18db) when you extract features using dump_data?

[hdmjdp]

@luvpine have you got good wav using 24k without modifying code?

[luvpine]

@hdmjdp No, I mean I'd like to make 24kHz wav file (high quality), not train using 24kHz data.

[hdmjdp]

I know, do you modify frame_size=160？

[belevtsoff]

@jmvalin Thanks a lot for the info! It would be great if you could give more details on what to change to switch to 22050Hz. Specifically, I was thinking of changing these things:

• increase NB_BANDS=24.
• change FRAME_SIZE to 256 (~11.6ms)
• OVERLAP_SIZE=256
• TRAINIG_OFFSET=128
• WINDOW_SIZE=512

I'm not sure about the pitch parameters though. In particular, currently:
PITCH_MIN_PERIOD 32
PITCH_MAX_PERIOD 256
PITCH_FRAME_SIZE 320

My guess is that the first two should be increased to match the difference in sampling rates, but should they be power of two? Also I guess the last parameter should be increased to match the window length. Do these changes sound right to you? Thanks a lot!

[luvpine]

@hdmjdp @belevtsoff In my case, change 16kHz -> 24kHz.

Key changing things:

• increase NB_BANDS=20
• change FRAME_SIZE = 480
• change total features 55 -> 59.

Also I changed nb_features , nb_total_features, sampling rate, feature index num(because features is increased), Conv1D layer filters size.

[belevtsoff]

@luvpine thanks! This set up sort of works, but I'm experiencing lowering of the pitch. Have you observed anything like this? (I've changed PITCH_FRAME_SIZE=480 as well)

[belevtsoff]

UPD: actually it keeps producing really bad results for me with 24kHz, can't figure out what the problem is.

[opencvbaby]

I tried 44.1KHz and 48KHz, also produced bad waveform, train and validation loss are all right, don't know why

[opencvbaby]

solved the problem, now 44.1KHz works well

[belevtsoff]

@opencvbaby Great! Can you share the full recipe?

[opencvbaby]

@belevtsoff Sorry, I can't share the recipe. There are some noise in high frequency, I am still optimizing it. Maybe you should take care of pitch_embedding part

[OswaldoBornemann]

@opencvbaby @luvpine could you share the parameter you have changed like nb_features , nb_total_features, sampling rate, feature index num and Conv1D layer filters size ? thanks a lot.

[linzai1992]

@belevtsoff Have you figure out your problems? Could you give some advice?

[xiaoyangnihao]

pitch_embedding

@opencvbaby hi, what do you mean by taking care of pitch_embedding part? can you make some explainations?