From Main to the search routine in Sphinx 3 s3accurate

1
From Main() to the search routine in Sphinx 3
(s3accurate)

• Arthur Chan
• July 8, 2004

2
This presentation

• Design goals of Sphinx 3 (s3accurate) (4 pages)
• How hard is to trace the Sphinx code? (2 pages)
• How to trace source code in general? (1 page)
• A tour from the main function to the search
  routine. (11 pages)
• Next time detail of s3accurate or we could start
  s3fast.

3
Design Goals of a Speech Recognizer/Trainer

• Different goals of building a speech recognition
  software
• Software Accuracy
• Software Speed
• Software Usability
• Code Readability Is it entertaining or easy to
  read?

4
If we optimize on only one aspect of Design goal

• Optimizing Accuracy
• Use 1000xRT techniques to optimize recognition
  rate.
• Optimizing Speed
• Cryptic coding makes programmer unable to
  understand whats going on.
• Optimizing Usability
• Allow gt 1000 options for users in the application
• But the code become extremely complex.

5
Conflicts between goals

• Accuracy vs Speed
• Usage of approximate search usually give faster
  decoding time but lower accuracy.
• Whole sentence decoding make sense for many
  techniques but cause a lot of delay.
• Usability vs Readability
• To make a software to be very usable, the code
  will be more complicated.
• Complicated code is generally hard to read.

6
Design goal of Sphinx 3 (s3accurate)

• Accuracy gt Readability gt Usability gt Speed
• Sphinx 3 the flat lexicon version,
• was mainly used for research purpose.
• very modular and easy to be changed by the
  researchers with experience in C
• Not many difficult code optimization could be
  found.

7
How hard is to trace code of Sphinx? (My view).

• As many programs
• Sphinx is just as a set of commands which do
  additions, subtractions, multiplications and
  divisions in a specific orders.
• So
• It is not difficult to understand the code if the
  underlining algorithm is understood.

8
Is there a lot of things to read?

• Number of lines of code of all .c and .h files.
• 36143
• Number of lines in Harry Potter novels
• Book 1 9000 lines
• Book 2 9000 lines
• Book 3 12000 lines
• Book 4 21000 lines
• Book 5 27000 lines
• So
• It is much shorter than the whole series of Harry
  Potter novels.
• Actually, it is not much to read.

9
How to read the source code in general?

• Several advices
• Jot notes on the dependency of the code
• (A file can be found in README.tracing can be
  found in s3fast and SphinxTrain. )
• Jot notes on how certain parts of the code work
• Something useful but not necessary An editor
  with program statistics and hyperlink of function
  definitions.
• Such as Microsoft Visual C (?)
• Or emacs (?) or vi (?)
• Grab a comfortable chair.
• Be patient.

10
A tour from the main function to the search
routine Overview

• Physical layout of the code
• s3decode-anytopo
• Brief tour of the programs
• Initialization
• Processing
• Post-processing.

11
Physical Layout of s3

• Getting the code from Sourceforge
• cvs -dpserveranonymous_at_cvs.sourceforge.net
  /cvsroot/cmusphinx/ co archive_s3
• archive_s3
• S3 ltlt THE ONE
• S3.0 lt-Why is it there?
• S3.2 lt- Legacy implementation s3fast wo
  live-mode decoder
• S3.3 lt- Legacy implementation s3fast w
  live-mode decoder

12
s3

• config/ lt- for configuration of make
• doc/ lt- documentation
• include/ lt- header files
• src/ lt- all the .c files.
• lib/ lt- where the library will be
• bin/ lt- where the binary will be.

13
Inside src/

• libio/ file IO functions
• libutil/ useful data structure
• libfbs/ source for all searching code
• libfeat/ code for feature extraction

14
Inside libfbs/

• Several files with main() include
• align-main.c -gt the entry point for s3align
• main.c -gt the entry point for s3decode-anytopo
• allphone-main.c -gt the entry point for s3allphone
• astar-main.c -gt the entry point for s3astar
• nbestrescore-main.c -gt the entry point for
  s3nbestrescore
• dag.c -gt the entry point for s3dag

15
Logical structure of the code

• The forward search
• 1, Do GMM computation for every senones
• 2, Do search for 1 frames.
• 3, Iterate until the end of the utterances.

16
main.c

• Pseudocode of top level
• 1, Read command line arguments
• cmd_ln_define
• cmdline_parse
• 2, Initialization
• Initialize log table lookup, logs3_init
• Initialize feature feat_init
• Initialize models models_init
• 3, Processing of all control files (also do
  decoding)
• Process_ctlfiles

17
process_ctlfiles

• Pseudocode
• 1, Read in more parameters
• 2, For every cepstral file,
• Run decoding (decode_utt)

18
decode_utt

• Run forward search fwdvit
• Dump hypothesis/lattice/statistics to the output

19
fwd_vit

• 1, Initialization of the feature vector
• 2, For every frame
• a, Computation of feature
• b, Computation of Gaussian distribution
• c, Do forward search for 1 frame

20
We are almost done !

• We went-through the key logic that leads us from
  the main() function to the GMM computation and
  search
• Next time
• Each component in more details.