This lists all know bugs subsequent to release FC2.0.14. For earlier releases contact Technical Support.
------------------------------------------------------------------------------------------------
When a trigger grammar allows “words before”, the target word may be preceded by other words. If the utterance does not contain the target word, T2SI may restart incorrectly, losing some data.
------------------------------------------------------------------------------------------------
The RSC-464 lacks port1. Code written for the RSC-4128 that
accesses port1 will cause unpredictable behavior when ported to the RSC-464.
The problem can’t be identified until the masked code is executed. The
fix is to add a “port1 present” bit to a library “magic
byte” and generate a warning if code attempts to access port1.
------------------------------------------------------------------------------------------------
RTLipSync does not work correctly if the user does not begin speaking quickly. If an initial silence period occurs, the lips do not respond because it becomes confused by any extraneous noise. The fix is to add code to ignore such noise.
------------------------------------------------------------------------------------------------
Natural Timeset listens for a period of silence at the end of the time utterance to know when the time set is finished. If the talker pauses too long during the timeset, the last part will be missed and the time will be wrong. The fix is to increase the amount of silence allowed during the timeset utterance.
------------------------------------------------------------------------------------------------
The FC libraries include technologies that are “lmodules”. Such modules are loaded into the final binary build only if they are invoked by the application. The Timeset grammar was added to the library as a “pmodule”, forcing it to be included in every build, whether used or not. The fix is to make the timeset grammar an “lmodule”.
The Phyton library
management tool can be used to change the type of the problem module in order
to make a temporary fix. Assuming the PATH variable includes the Phyton bin directory do the following:
mclib -L fc4128.mcl AMPMFORCE_RSCGRAM_TIMESET_EN_US
------------------------------------------------------------------------------------------------
The FC libraries include technologies that are “lmodules”. Such modules are loaded into the final binary build only if they are invoked by the application. The Timeset net was added to the library as a “pmodule”, forcing it to be included in every build, whether used or not. The fix is to make the timeset net an “lmodule”.
The Phyton library
management tool can be used to change the type of the problem module in order
to make a temporary fix. Assuming the PATH variable includes the Phyton bin directory do the following:
mclib -L fc4128.mcl _NN_TIMESET_EN_US_MOD
------------------------------------------------------------------------------------------------
The RSC-4128 cannot run fast enough to keep up with the speech during TimeSet. It “catches up” after speech has finished, then reports what was said. This latency can be decreased by running with zero wait states. In some releases the latency is longer than it should be because extra unneeded processing is done. The fix is to reduce the amount of processing.
------------------------------------------------------------------------------------------------
The RS232 Tx pin has unrequested debug output present during Timeset. This may cause problems in applications that use the RS232 Tx pin for other purposes.
------------------------------------------------------------------------------------------------
The beat detect and peak detect technologies change the extAdd register, but do not restore it. This causes a problem if the application also explicitly uses the extAdd register. The fix is to restore the extAdd register in these technologies.
------------------------------------------------------------------------------------------------
Lipsync technology uses encoded data in the speech file that describes the mouth position and duration. If the duration values are too short, the mouth does not synchronizes properly with the speech. The fix is to modify the technology to deal with short durations.
------------------------------------------------------------------------------------------------
The SX “sentence” technology allows playing a sequence of words. When some of the words are encoded using ADPCM and others are encoded using SX6, some initialization may not be properly done, producing incorrect speech. The fix is to perform the correct initialization.
------------------------------------------------------------------------------------------------
SD/SV templates may be trained using a method that requires each utterance to be spoken only once (“1-pass”). If the background noise level is too high while training, the template cannot be made.
------------------------------------------------------------------------------------------------
The RSC-464 library provides “powerdown” code and sources for low-power states. Modifications to this code may cause unpredictable behavior upon wakeup. The fix is to strongly encourage developers always to use the library code functions and macros. In addition, the library code was made robust against this problem.
NOTE: This problem applies to the RSC-464 only, but Sensory recommends using only the library powerdown routines on the RSC-4128.
------------------------------------------------------------------------------------------------
A T2SISD trigger can use both SI and SD words. When a NOTA (non-recognized utterance) is heard, T2SI re-initializes for another listening attempt, but does not correctly initialize the SI portion. So the SD word may be recognized, but the SI word is not recognized, even on repeated attempts. A silence interval of about 6 seconds may restore recognition of the SI word.
------------------------------------------------------------------------------------------------
The FC library defines T2SI_DEFAULT_TIMEOUT, which may be used for trigger or command vocabularies. Internally the library uses this flag value to set the actual timeout to either 3 seconds (command) or 60 seconds (trigger). The default trigger timeout value of 60 is incorrect and should be 0 (never timeout).
------------------------------------------------------------------------------------------------
The FC libraries include technologies that are “lmodules”. Such modules are loaded into the final binary build only if they are invoked by the application. A number of diagnostic/rs232 utility routines were added to the MC464.MCL library as “pmodules”, forcing them to be included in every build, whether used or not. The fix is to make these routines “lmodules”.
NOTE: This affects only RSC-464 applications.
The Phyton library
management tool can be used to change the type of the problem module in order
to make a temporary fix. Assuming the PATH variable includes the Phyton bin directory do the following:
mclib -L fc464.mcl DELAY232_MOD DIAG232D16_MOD DIAG232D8_MOD mclib -L
fc464.mcl DIAGCRLF_MOD DIAGH16VAL_MOD DIAGH16_MOD
mclib -L fc464.mcl DIAGH24_MOD DIAGH32_MOD DIAGH4_MOD
mclib -L fc464.mcl DIAGH8_MOD PUTDIAG232_MOD
mclib -L fc464.mcl PUTHEXDIGIT232_MOD PUTHEXDIGIT232_MOD
NOTE: Many of these functions are
called internally by Sensory technologies, so making them lmodules
will not necessarily remove them from the binary.
------------------------------------------------------------------------------------------------
The duration of a wordspot (SDWS, SVWS) utterance should be 2-6 syllables for best results. The technology allows longer utterances with reduced accuracy, failing somewhat gracefully as the duration increases. Due to a bug, if the duration of the training utterance is longer than roughly one second, then training will succeed but the WS template can never be recognized.
------------------------------------------------------------------------------------------------
This intermittent problem applies to SD/SDF only, not SV and not WS. When a “large” SD set is recognized (more than 3 templates, but more-or-less depending on the knob setting), recognition is done in two passes: the first pass screens all templates to select the best few for further processing; the second pass does a finer analysis of the selected candidates. When running at zero wait-states, the initial screening can make errors, which can occasionally result in not submitting the correct candidate template for the second pass, causing recognition to fail.
------------------------------------------------------------------------------------------------
When a T2SI vocabulary becomes too big to fit in memory, QT2SI employs a method of re-using certain resources. The recognition code had a bug that caused some utterances not to be recognized in this case.
------------------------------------------------------------------------------------------------
Date Last Modified: 2009-July-15