Call Recording

807 views
Skip to first unread message

lucif...@gmail.com

unread,
Apr 11, 2019, 4:45:17 AM4/11/19
to kaldi-help

I want to train a model to recognize the voice,over call recording .

(1)So I recorded the voice using microphone and on call .

(2)I trained the model for both the cases.
But It is performing bad in on call recording case.

(3)So I want to know is there any sampling rate difference between on call recording and microphone recording.
Does just merely overlaying call noise on microphone recording-->acts as call recodring?

Daniel Povey

unread,
Apr 11, 2019, 4:11:59 PM4/11/19
to kaldi-help
If there was a sampling rate difference -- and that obviously depends on what sample rate you recorded the signals at --  the scripts would have complained.  It may just be that the telephone audio is lower quality.


--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To post to this group, send email to kaldi...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/4cc20bf4-b666-4033-a212-abfcd99d9162%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Jeff Brower

unread,
Apr 11, 2019, 4:55:48 PM4/11/19
to kaldi-help
Lucifer-

Send me short wav files from both your direct mic and on-call recordings (same speaker segment, preferably aligned within 100 msec), and I will verify your sampling rates and run 2-D Spectrograph analysis to see what differences might be causing the issue.  It would look something like this:

  https://www.signalogic.com/images/EVS_decoder_NB_filter_Fs_conv_vs_external_Fs.png

The above example compares the 3GPP EVS decoder built-in NB output with its WB output down-sampled by an external algorithm (just an example).  We have to do this type of comparison analysis frequently for our customers.  Not just ASR training that can be very sensitive to slight differences in the frequency domain.

-Jeff

lucif...@gmail.com

unread,
Apr 12, 2019, 1:02:33 AM4/12/19
to kaldi-help
I have read this 
"Music signals are typically sampled at 44,100 Hz (or 44,100 samples per second). Due to the Nyquist theorem, this means that audio with frequencies of up to 22,050 Hz can be faithfully captured by sampling. Speech signals have less high frequency (only up to 8000 Hz) information so a sampling rate of 16,000 Hz is typically used. Speech over conventional telephone lines and most mobile phones is band-limited to about 3400 Hz, so a sampling rate of 8000 Hz is typically used for telephone speech."

(1) While trainning by gmm-hmm model I down-sampled every audio file using --allow-downsample =true to (8000 hz) and 
(2) and my call Recorder app recorded audio file at 44100 HZ
(3)So Is it the problem of sampling rate(band limit as mentioned above)or the audio quality of telephone?


On Friday, April 12, 2019 at 1:41:59 AM UTC+5:30, Dan Povey wrote:
If there was a sampling rate difference -- and that obviously depends on what sample rate you recorded the signals at --  the scripts would have complained.  It may just be that the telephone audio is lower quality.


On Wed, Apr 10, 2019 at 10:45 PM <lucif...@gmail.com> wrote:

I want to train a model to recognize the voice,over call recording .

(1)So I recorded the voice using microphone and on call .

(2)I trained the model for both the cases.
But It is performing bad in on call recording case.

(3)So I want to know is there any sampling rate difference between on call recording and microphone recording.
Does just merely overlaying call noise on microphone recording-->acts as call recodring?

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.

Daniel Povey

unread,
Apr 12, 2019, 2:03:14 AM4/12/19
to kaldi-help
It's not an issue of sampling rate

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.

To post to this group, send email to kaldi...@googlegroups.com.
Reply all
Reply to author
Forward
0 new messages