[hts-users:04213] Re: endianness

[Keiichiro Oura <uratec@sp.nitech.ac.jp>]

Hi,

HTS and demo are not checked with big endian machine.
Does anyone know?

Regards,
Keiichiro Oura


2015-01-29 10:52 GMT+09:00 Xingyu Na <asr.naxingyu@gmail.com>:
> I think you are right about this. I remember in early versions of HTS demos,
> there is a BYTESWAP option in configuration and also tunable in the Makefile
> of data and Config.pm. I don't know if they have removed since 2.3.
> Perhaps Keiichiro could explain.
>
> Xingyu
>
>
> On 01/29/2015 09:34 AM, Matt Shannon wrote:
>>
>> Hi Xingyu,
>>
>> The failures I'm talking about would happen when converting between audio
>> data and speech parameters rather than at the modelling stage. For example
>> in data/Makefile.in several times x2x is called on the raw audio data, which
>> will produce nonsensical results on a big endian machine if I'm right about
>> SPTK using native endian by default.  Also in scripts/Training.pl, raw2wav
>> would fail to produce a valid wav file when not using STRAIGHT (since wav
>> files should always be little endian but raw2wav does no conversion unless
>> you explicitly tell it to) or MATLAB would fail to read in a valid spectrum
>> if using STRAIGHT (since "ieee-le" is explicitly specified).  I presume
>> support for big endian machines has just been dropped and that seems
>> reasonable enough, but I just wanted to check!
>>
>> Matt
>>
>>
>> On 29/01/15 01:25, Xingyu Na wrote:
>>>
>>> Hi Matt,
>>>
>>> I remember there are NATURALREADORDER and NATURALWRITEORDER options in
>>> HTK. Can't that help?
>>>
>>> Best,
>>> Xingyu
>>>
>>> On 01/29/2015 09:19 AM, Matt Shannon wrote:
>>>>
>>>> Hi all,
>>>>
>>>> I was wondering recently about whether the HTS demo supports CPU
>>>> architectures which are native big endian.  As far as I can ascertain,
>>>> all SPTK tools use native endian by default.  From looking at the HTS
>>>> demo scripts, it looks like they will not work on a big endian machine
>>>> either at training time or at synthesis time (but it's hard to know
>>>> for sure without access to a big endian machine!)  I know in older
>>>> versions there was a BYTESWAP variable that got set at configure
>>>> time.  I presume it was decided at some point not to support big
>>>> endian machines?
>>>>
>>>> Thanks,
>>>>
>>>> Matt
>>>>
>>>>
>>>
>>>
>>
>
>