[hts-users:03196] feature extracting problem
- Subject: [hts-users:03196] feature extracting problem
- From: li jay <lij.acd@xxxxxxxxx>
- Date: Fri, 9 Mar 2012 05:16:39 +0800
- Delivered-to: hts-users@xxxxxxxxxxxxxxx
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=20120113; h=mime-version:date:message-id:subject:from:to:content-type; bh=KJ1W2Nv/baUCgoXeMB4wRTZlAOJ9zvKWrDylz3AIAZU=; b=qHhnDmmYRFwpuqTiL13trEvrnQgzQyhb3u9BWWEeiNlMBALV+erZCe5Gx7Hv46dQP4 g4doCVI+7wK/szRjAsl1zVyPOKBw2tj4Sq5aujDEkEAzSOZDJ5N+c8UWVCVP9gvLYjfF /hxmADXL+0oDY26ho/Pqa9v9QUI4FW19X+KXEOXb6SMQwzfK/TRCEucNIH7AyMWQ5SsR aLj6YuLZ32gdyy4K/nNjgkGj9bPvdazHh8BvDY1QrXFObfH/dDdgHgdAbmlMJw+rmy2Z vaiK+TJT6Z/mmVhUBI3ce5FatwTLs96ruc7J+Akpupnh4vk+0J3TWMDr+gBY6JYbqk6t DTww==
Hi,
I'm using HTS-2.1.1 and SPTK-3.3. I've trained average speaker models to adapt to a speaker's voice. However there were strong noise in the adapted voice, though there were very little noice in the average voice. I thought it was resulted from training data containing noise. I denoised wav files in the training data. But I encountered the problem I've never seen before the denoising. When I was extracting LSP features using the folloing command, there was a error said:
/usr/local/SPTK/bin/mgcep -a 0.0 -c 1 -m 39 -l 512 -o 4 | /usr/local/SPTK/bin/lpc2lsp -m 39 -s 16 -l -n 512 -p 8 -d 1e-6 > mgc/spkr004_sen0032.mgc
theq() : determinant of the normal matrix is too small!
I searched on the web, and I found make the mimimum value of the determinant of the normal matrix lower could be helpful. So I used the command below:
/usr/local/SPTK/bin/mgcep -f 0.00000000001 -a 0.0 -c 1 -m 39 -l 512 -o 4 | /usr/local/SPTK/bin/lpc2lsp -m 39 -s 16 -l -n 512 -p 8 -d 1e-6 > mgc/spkr004_sen0032.mgc
theq() : determinant of the normal matrix is too small!
But the error still appeared.
There was a thread in the archive said "SPTK-3.3 support 32-bit machine only." However I can extract features well previously and I'm using 64-bit machine system. After denoising I listened to the denoised noise, and found the noise was removed effectively. So I think there is no problem on denoising. Could you give me some advice or opinion on this problem?
Regards,
Jay
- Follow-Ups
-
- [hts-users:03198] Re: feature extracting problem, Keiichiro Oura