[Subject Prev][Subject Next][Thread Prev][Thread Next][Date Index][Thread Index]

[hts-users:04543] Data normalization

I saw a few posts on the mailing-list about data normalization, I would like to know what are the best practices

I’m starting with 16kHz samples that I upsampled to 48kHz using SoX, which seems to renormalize the gain and cause clipping later

Right now I’m not using wav2raw to convert the upsampled wav to raw as this was causing some issues with clipping, but I’m using sox in this way:

sox $input -t raw $output

after this I still had many files skipped during feature building because sox would renormalize some values to use the whole amplitude from -32768 to 32767 which would make the test skip it:

if [ -s $< -a $${min} -gt -32768 -a $${max} -lt 32767 ]

so I added a gain normalization which helps with this

sox $input —norm=-5 -t raw $output

but I still get a lot of clipping warning during the synthesis phase after training

Are there suggestions on a proper way to normalize data, both which tools to use (sox or wav2raw) and which options or target gain?

Should we also update the check rule to

if [ -s $< -a $${min} -ge -32768 -a $${max} -le 32767 ]

( <= and not <) to allow for all normal short values, or is there a reason it was using -gt and not -ge

Dr Maël Primet, PhD    
Founder & CTO  
+33 (0) 6 51 53 38 27
Skype: maelpr
Twitter: @mael_p

[hts-users:04544] Re: Data normalization, Simon King