blob: 1bfec160572cee5c0cd8d20c1c57e7a58a84e842 (
plain) (
tree)
|
|
nerv.speech_utils = {}
function nerv.speech_utils.global_transf(feat_utter, global_transf,
frm_ext, frm_trim, gconf)
-- prepare for transf
local input = {feat_utter}
local output = {feat_utter:create()}
-- do transf
global_transf:init(input[1]:nrow())
global_transf:propagate(input, output)
-- trim frames
expanded = gconf.cumat_type(output[1]:nrow() - frm_trim * 2, output[1]:ncol())
expanded:copy_fromd(output[1], frm_trim, feat_utter:nrow() - frm_trim)
collectgarbage("collect")
return expanded
end
function nerv.speech_utils.feat_expand(feat_utter, frm_ext, gconf)
local rearranged
if frm_ext > 0 then
local step = frm_ext * 2 + 1
-- expand the feature
local expanded = gconf.mmat_type(feat_utter:nrow(), feat_utter:ncol() * step)
expanded:expand_frm(feat_utter, frm_ext)
-- rearrange the feature (``transpose'' operation in TNet)
rearranged = gconf.mmat_type(feat_utter:nrow() - frm_ext*2, feat_utter:ncol() * step)
rearranged:rearrange_frm(expanded, step, frm_ext, feat_utter:nrow() - frm_ext)
else
rearranged = feat_utter
end
collectgarbage("collect")
return rearranged
end
function nerv.speech_utils.normalize(mat, global_transf)
-- prepare for transf
local input = {mat}
local output = {mat:create()}
-- do transf
global_transf:init(input[1]:nrow())
global_transf:propagate(input, output)
return output[1]
end
|