PocketSphinx and Sphinx-4

[creative64]

Hi,

I'm new to Sphinx and need help in some clarifications regarding PocketSphinx
and Sphinx-4.

1. Does PocketSphinx support continuous HMMs or it only supports SCHMM (Semi Continuous HMMs) ?

2. Does PocketSphinx also support Floating Point arithmetic ?

3. If the answer to 02 is yes then what would typically be accuracy difference between FixedPoint and FloatingPoint versions of PocketSphinx for a typical memdium size (say 1000 word) vocabulary application ?

4. Does PocketSphinx support FSN style of grammar ?

5. Does Sphinx-4 support FSN style of grammar ?

6. What is the fundamental technical difference between PocketSphinx and Sphinx-4 that contributes to inaccuray of PocketSphinx compared to Sphinx-4 and what would be typical accuracy degradation ?

Thanks.

[Nickolay V. Shmyrev]

1. Does PocketSphinx support continuous HMMs or it only supports SCHMM
   (Semi Continuous HMMs) ?

Pocketsphinx supports continuous HMMs

Does PocketSphinx also support Floating Point arithmetic ?

Pocketsphinx supports FP arithmetics

If the answer to 02 is yes then what would typically be accuracy difference
between FixedPoint and FloatingPoint versions of PocketSphinx for a typical
memdium size (say 1000 word) vocabulary application ?

If you'll compare them on semicontinuous models the difference would be minor
because of rounding errors. Usually the choice between fixed and floating is
done by presence of FPU unit in CPU.

Does PocketSphinx support FSN style of grammar ?

Yes

Does Sphinx-4 support FSN style of grammar ?

Yes

What is the fundamental technical difference between PocketSphinx and
Sphinx-4 that contributes to inaccuray of PocketSphinx compared to Sphinx-4
and what would be typical accuracy degradation ?

Inaccuracy of pocketsphinx comparing to sphinx4 is a myth. Actually
pocketsphinx is more accurate because it decodes in multiple passes.

[creative64]

Thanks a lot for your comments nshmyrev. Couple more related questions.

If you'll compare them on semicontinuous models the difference would be
minor because of rounding errors. Usually the >>choice between fixed and
floating is done by presence of FPU unit in CPU

1. How about difference in fixed point and floating point for continuous HMM implementation.

Inaccuracy of pocketsphinx comparing to sphinx4 is a myth. Actually
pocketsphinx is more accurate because it decodes in >>multiple passes

1. What then would be the prime reason for somebody to prefer Sphinx-4 over PocketSphinx ?

2. Does Sphinx-4 too support fixed point arithmetic ?

Thanks.

[Nickolay V. Shmyrev]

How about difference in fixed point and floating point for continuous HMM
implementation.

On an4 on usual intel machine fixed point is 16% WER, floating 16.2% WER.
Fixed point is way slower because CPU is not used. When vocabulary is bigger,
fixed point is a bit less precise, but the difference is minor comparing to
other factors.

What then would be the prime reason for somebody to prefer Sphinx-4 over
PocketSphinx

If you design large scale system flexibility is more important than speed. In
this case Java wins. ASR applications is not just engine, it requires
integration with various modules from job scheduling to user database query.

Does Sphinx-4 too support fixed point arithmetic

No

[creative64]

Thanks a lot nshmyrev.