todo

ears-0.32, linux下有用的语音信号处理工具包

=====| soon |=============================================================
- rec. files should be named <basename>.<recognizer>.<feature extractor>
  filename entries can then be also without path.
- "There are new words" + ESC --> bailout
- bailing out of UI_RAW binaries should first send newline
- if no words are trained with UI_RAW: listen says bogus file
- implement flag for train_ears to process WAVs directly into patterns
- use util/configure to check for existence of OSS sound driver, compile
  both into the apps, choose at runtime depending on /dev/sndstat
  new config option: sound driver
- use util/configure to check for existence of UI libraries, compile all
  into the apps, choose at runtime depending on pr_config::check_libs()
  new config option: user interface, remove NO_CURSES
- why is libm linked in with elf version???
- implement test flag (Niels)
- check all comments for kludges
- listen-recw w/ UI_NCURSES has a hole in the border (pr_listen.cc empty; )
- use two different protocols for CL and NCL versions (yarec ui code)
- "Do you want to use the digits..." + NO with other than default as
  basename, should we really quit?
- further fine-tune 16-bit endpoint detection
- SunSound::save_sample() isn't implemented

**********C++ / OO design issues:
- draw Booch diagram of application
- rename TrainEars-/ListenProtocol to TrainEars-/ListenApplication
- initialization in Oss/Voxwaresound could be handled by device classes
- modularize feature
- modularize endpointer
- implement abstract ui event handler (callback)
- make a ConfigP base class with diff. derived classes for train_ears, listen
- screen::close() is called twice on a normal exit
- remove class messages entirely from pr_config.h
- words should be a parameter to training
- singletons should be implemented otherwise, no delete...new in protocols
  by using AppResourcesList or giving them as ref parameters to others
- write private empty copy ctors/ass op's for classes that aren't copied
- func try_to_open(ears_file) for words and other files
- there are still too much dependencies on ears/*.h, esp. in modules!
- take yarec/modules/sample.h and make ears/sample a has-a sample class

**********standard C++ library usage:
- use vector<>/valarray<> for arrays (valarray needs gcc-2.8.0)
- use vector<Word*> instead of list<Word*>, derive words from it publically
  all other members should be private
- use random_shuffle to find words
- use stream_iterators throughout the code: ostream_iterator<X> oit(cout," ");
  copy (i,j,oit);

**********exception handling:
- implement real exception handling (gcc-2.8.0)
- exceptions in speechstream.cc and feature.cc
- exceptions in cursesd.*
- what do we do if a new version introduces new cfg[]s that aren't in
  the existing earsrc.  
  exception handling?  simply a warning?  version option?

==| not so soon |=========================================================
- include new version of rasta, don't use it as library, so people can
  build from the distribution themselves
- modal vocabularies to keep active word list small (M. Ward)
- implement 'Calibrate mic' window for making endpoint detection more robust
- in listen: the possibility to choose from a list of basenames
- check endianness when writing/reading datafiles
- write README section 'What if recognition is unsatisfying?'
- do not train recognizer if words are incomplete
- support /dev/random in myrandom.h
- check existence of all patterns given in recognizer file
- memory leaks in Regex and ncurses, NDTW::ref.patterns aren't freed as
  well as some things in libmrasta.
- play with /proc/cpuinfo on Linux
- speedup: calculate features while recording (enough soundcard buffer?)
           (new option PROCESSING ONLINE/OFFLINE)
           calc DTW while recording!!!
- implement option to hardcode noise level
- automagically always record noise sample(s) for endpoint detection
  maybe a small recurrent net could learn to choose between speech/noise?
- use the EBUSY return value from /dev/dsp
- ioctl(fd,SOUND_PCM_SUBDIVIDE,&div) with div<5 decreases sound buffer
- "SOUND_BITS/SPEED doesn't match the recorded patterns!  Proceed?"
- "Changing SOUND_BITS/SPEED when not having recorded all patterns
   might yield unexpected results.  Proceed?"
- pattern::read(): ask user if bogus pattern files should be removed
- implement another good, fast recognizer
- DTW: implement a VQ preprocessor (see Rabiner)
- implement robust training (see Rabiner)
- does the rasta vector include energy information?
- implement hybrid approach to endpoint detection (see Rabiner)
- automagically always record noise sample(s) for training
- implement handling of contradictory recognition results (status bar etc.)
- save word.actions in net file - no more need to check for words
- implement Feature=NONE (for recognizers that can/must digest raw data)
- should we switch to lowercase option names?
- during training, output how much words still to speak, percentage
- check config options for legality
- "There are xx samples missing.  Do you want to record some of them?"
- write more words files as examples
- make word list a map 
- remove all stdio usage, also in cursesw.* 
- include Tilman's German message catalog
= implement ears
- ears: should we renice the spawned shell for recog performance?
= implement listen fully
= test ears
= implement cross validation program
- write texinfo
- include Hermansky paper in doc/
- say "Huh?" over the speaker when recognition fails  8-)