Tball Literacy Assessment
meeting notes |
Abe's Webpage for Tball Related Research
This page is for resources related to the tball project. Please use the links above to search within this page for your convenience.
The public tball web page is here.
... as of July (2007)
... as of May 23,
- Posted the slides for the tball retreat.
... as of May 5,
... as of Apr 3
... as of Mar 2
- I make a repository for research logs, mainly for my own record,
but also to share info if need be. See them here:tball blogs
... as of Feb 27
- I posted the fixed and cmu versions of all the transcriptions here
- I also posted the script I used to make it these different versions
... as of Feb 21
... as of Feb 6
... as of Feb 5
... as of Jan 30
- I posted alphabet task audio files from May
- and the corresponding transcriptions here
... as of Jan 8
- I posted meeting notes for the usc asr team here.
... as of Dec 2
... as of Sept 23
... as of aug 21
... as of Aug 14
- I gathered some stats
about the transcriptions...it should help
to see where we have enough transcriptions and where they are needed.
... as of July 25
- I uploaded the Eurospeech poster draft (in a single poster format).
...as of June 7
- I posted more sentences (current total 626 transcriptions)
- I posted some of the kwordlist done so far (923).
- I made a counter to
automatically keep track of the transcription progress.
...as of June 3
- I posted some preliminary sentence transcriptions that Daylen did.
- I posted the kindergarten wordlist (downsampled to 16k)
...as of may 2
I set up an online message board for use in discussion,
brainstorming, collaboration, etc. Because it is pretty insecure
(unsecure), its not a good idea to have external links to it. Email
me if you don't know where it is or if you need help. Right now
there are discussion about the db, asr, the recordings, and a main
page that everyone should sign to get a feel for how easy it is.
I posted the child table from the DB in text format, for relating
the child info to the child id from the wavefile names...I think Mateo
is making different acoustic models for different ages and genders.
...as of apr 22
- I manually corrected some forced alignments to try with HInit/HRest (posted under transcriptions)
...as of apr 20
- Updated the dictionaries with the new transcribed data (6747 transcriptions total)
- I updated the scripts for converting fixing transcriber errors and converting to cmu format.
- I converted all the transcriptions so far to cmu format and zipped them up for Jorge
...as of mar 23
- g1 wordlist transcriptions are posted
- Sentence audio files are posted--downsampled to 16k and zipped (there still seems to be a problem with the headers--the files that got screwed up have 1kb)
I fixed a problem with the dictionaries the database there
are numbers listed as words (when in a word list) and listed as
numbers (..in the number lists). Before in the dictionaries, they
were separate as well, whereas now they are all mapped to
- I've uploaded some scripts that I've used in making the
dictionaries and processing the transcriptions. They may prove to
be useful and time-saving, or not (usually they aren't well
commented)... Please ask me if you have Q's
- I made a recognizer based on the transcriptions so far. It
seems to work fairly well (~65%) when evaluated on the training
live it even works sometimes on my voice when I talk like an
elementary-schooler. The segmentation from the decoding seems to be
good on the word level. I put the models and word lists below under
ASR. I still have yet to implement the tied-state "sp" and
- cmu_lookup.pl: This takes
a word as the commandline arg and returns the pronunciation(s)
This is a routine which is meant to be included in other programs in
order to convert the tball transcriptions to cmu symbols
This is a routine that fixes common errors in transcriptions so you
won't get an out of vocab error on any of the transcription
This program uses the above scripts to compile a dictionary. The
directory containing the transcriptions is specified on the command
line and options are set using flags.
- play.tcl:This is a little script
to emulate the SoX "play" command for if I want to play a wav file
from the command line when I am not using linux (requires snack library).
a program for finding the
minimum edit distance between two strings
- script for analysing duration
of segmented words (work in progress)
- sox downsampling exammple
- I posted the script I used to translate between transcription formats
- training script
- monophone list
HVite.exe -H hmm2/macros -H hmm2/hmmdefs -w wordnet -S train.scp -l labs -p 0.0 -s 5.0 dict monophones1 to decode the already transcribed data.
- with an updated config file, run
HVite.exe -H hmm2/macros -H hmm2/hmmdefs -C config -w wordnet -p 0.0 -s 5.0 dict monophones1 to run the recognizer live.
Papers and Presentations
Viterbi School of Engineering
Last modified: Thu Jan 10 21:43:43 PST 2008