require 'libhtkio' require 'speech_utils' local TNetReader = nerv.class("nerv.TNetReader", "nerv.DataReader") function TNetReader:__init(global_conf, reader_conf) self.feat_id = reader_conf.id self.frm_ext = reader_conf.frm_ext self.gconf = global_conf self.global_transf = reader_conf.global_transf self.debug = global_conf.debug if self.debug == nil then self.debug = false end self.feat_repo = nerv.TNetFeatureRepo(reader_conf.scp_file, reader_conf.conf_file, reader_conf.frm_ext) self.lab_repo = {} for id, mlf_spec in pairs(reader_conf.mlfs) do self.lab_repo[id] = nerv.TNetLabelRepo(mlf_spec.file, mlf_spec.format, mlf_spec.format_arg, mlf_spec.dir, mlf_spec.ext) end end function TNetReader:get_data() if self.feat_repo:is_end() then return nil end local res = {} local frm_ext = self.frm_ext local step = frm_ext * 2 + 1 -- read HTK feature local feat_utter = self.feat_repo:cur_utter(self.debug) -- global transf local transformed = nerv.speech_utils.global_transf(feat_utter, self.global_transf, frm_ext, self.gconf) res[self.feat_id] = transformed -- add corresponding labels for id, repo in pairs(self.lab_repo) do local lab_utter = repo:get_utter(self.feat_repo, transformed:nrow(), self.debug) res[id] = lab_utter end -- move the pointer to next self.feat_repo:next() collectgarbage("collect") return res end