[hts-users:02766] the increasing tendency of MCD curves
- Subject: [hts-users:02766] the increasing tendency of MCD curves
- From: Hui LIANG <tshlmail-hts@xxxxxxxxx>
- Date: Mon, 11 Apr 2011 08:07:34 -0700 (PDT)
- Delivered-to: hts-users@xxxxxxxxxxxxxxx
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=yahoo.com; s=s1024; t=1302534454; bh=sNs4ih/TkYCDwrXo4ZOuQ+4BRP3gE8F7R9EaQlrgAoM=; h=Message-ID:X-YMail-OSG:Received:X-Mailer:Date:From:Reply-To:Subject:To:MIME-Version:Content-Type; b=E7Yt4sEupHbPbXdEAX6eCB6aGZuNXyIpR8QhHQBUJHEGKFtjHZ5wp1jtlNo3q9DwiZlgQLFC9c780Prqv1o59yUAAChhPLMGIiCmoZAgBLvyVCgb1Py8xNi0yuGFbgRmb0ANL+RZwG41ugsrf0u1gZiFoMDPaPU4hBiKtB4GyhU=
- Domainkey-signature: a=rsa-sha1; q=dns; c=nofws; s=s1024; d=yahoo.com; h=Message-ID:X-YMail-OSG:Received:X-Mailer:Date:From:Reply-To:Subject:To:MIME-Version:Content-Type; b=NyTzeQZdjMg+i9RFtFjduEYGBIGEmypXdKPKc7g/LBK/TiElTYcZWHHB2xt7woNNtid8mqNb9AIfW0v5glYgcQ1wr72wTNe08QCrCzMchtwYz1W55M9A29swtJSo5Y1sCfR4ofX/yMpQKpe2ZPXJXazQQgNNN92q7y0lrmhAbGY=;
Hello everybody,
I happened to find a phenomenon about which I personally felt kind of interesting. I was adapting an English average voice model with English adaptation data), using the CSMAPLR algorithm in HTS 2.1. As the attached figures show, more iterations of transform estimation (slightly) increased mel-cepstral distortion of English testing data, while I expected the opposite. All the four target speakers had reasonably natural English accents, J being a native speaker.
I'd like to know if anyone has observed the same phenomenon before or has any comments on these curves, though I understand that non-MGE (minimum generation error) criteria for transform estimation don't guarantee reduction of mel-cepstral distortion.
Thanks a lot!
Hui LIANG
Attachment:
MCD_intra.pdf
Description: Adobe PDF document