Help for Sensory FluentChip FC5 Software Development  Kits

 

 

This lists all know bugs subsequent to release FC5.0.6.

 

 

1.     T2SI recognition works erratically when large acoustic models are in expansion memory.

2.     T2SI recognition works poorly when large acoustic models are in external memory.

3.     T2SI recognition debug reports occasional erratic powers

4.     T2SI recognition makes strange mistakes with large vocabularies when the acoustic model is in internal ROM code space.

 

------------------------------------------------------------------------------------------------


 

T2SI recognition works erratically when large acoustic models are in expansion OTP memory.

 

Summary

The T2SI acoustic model data can be located onchip in either code space or const space, or in external memory. For certain large vocabularies the const space (expansion OTP memory) address for portions of the acoustic model is calculated incorrectly. This bug was introduced by the fix for net problems in external memory.

 

First Release Affected: FC5.0.9

Last Release Affected: FC5.0.9

Severity: MEDIUM (non-fatal but apparent effects)

Workarounds: none

Recommended Solution: upgrade to FC5.0.10 or later.

 

 

 


 

T2SI recognition works poorly when large acoustic models are in external memory.

 

Summary

The T2SI acoustic model data can be located onchip in either code space or const space, or in external memory. For certain large vocabularies the external address for portions of the acoustic model is calculated incorrectly.

First Release Affected: FC5.0.8

Last Release Affected: FC5.0.8

Severity: MEDIUM (non-fatal but apparent effects)

Workarounds: locate the acoustic model onchip if there is room.

Recommended Solution: upgrade to FC5.0.10 or later.

 

 

 

------------------------------------------------------------------------------------------------


 

T2SI recognition debug reports occasional erratic powers

 

Summary

The debug output for T2SI contains utterance and silence level powers. Occasionally these values are significantly too big or too small, reflecting bugs in the underlying FFT and saturation logic. These bugs could also potentially reduce recognition accuracy. Any accuracy reduction would be slightly and probably not noticeably.

First Release Affected: FC5.0.6

Last Release Affected: FC5.0.7

Severity: MEDIUM (non-fatal but apparent effects)

Workarounds: none

Recommended Solution: upgrade to FC5.0.8 or later.

 

 

 

------------------------------------------------------------------------------------------------


 

T2SI recognition makes strange mistakes with large vocabularies when the acoustic model is in internal ROM code space.

 

Summary

A T2SI grammar for a large vocabulary may require more than 255 phonemes. When the acoustic model is in internal ROM code space, computation of probabilities for phonemes numbered higher than 255 is wrong (the computation is correct for smaller acoustic models in internal ROM code space and for all acoustic models in external parallel flash). This error may cause recognition to fail on of some parts of the vocabulary, while other parts are recognized normally.

First Release Affected: FC5.0.6

Last Release Affected: FC5.0.7

Severity: MEDIUM-HIGH (product operation may be impaired)

Workarounds: none

Recommended Solution: upgrade to FC5.0.10 or later.

 

 

------------------------------------------------------------------------------------------------

 

Date Last Modified: 2010-Jan-04