<s>
Speech	B-Application
recognition	I-Application
is	O
an	O
interdisciplinary	O
subfield	O
of	O
computer	B-General_Concept
science	I-General_Concept
and	O
computational	O
linguistics	O
that	O
develops	O
methodologies	O
and	O
technologies	O
that	O
enable	O
the	O
recognition	O
and	O
translation	O
of	O
spoken	O
language	O
into	O
text	O
by	O
computers	O
with	O
the	O
main	O
benefit	O
of	O
searchability	B-Algorithm
.	O
</s>
<s>
It	O
is	O
also	O
known	O
as	O
automatic	B-Application
speech	I-Application
recognition	I-Application
(	O
ASR	O
)	O
,	O
computer	B-Application
speech	I-Application
recognition	I-Application
or	O
speech	B-Application
to	I-Application
text	I-Application
(	O
STT	O
)	O
.	O
</s>
<s>
It	O
incorporates	O
knowledge	O
and	O
research	O
in	O
the	O
computer	B-General_Concept
science	I-General_Concept
,	O
linguistics	O
and	O
computer	O
engineering	O
fields	O
.	O
</s>
<s>
The	O
reverse	O
process	O
is	O
speech	B-Application
synthesis	I-Application
.	O
</s>
<s>
Some	O
speech	B-Application
recognition	I-Application
systems	O
require	O
"	O
training	O
"	O
(	O
also	O
called	O
"	O
enrollment	O
"	O
)	O
where	O
an	O
individual	O
speaker	O
reads	O
text	O
or	O
isolated	O
vocabulary	O
into	O
the	O
system	O
.	O
</s>
<s>
Speech	B-Application
recognition	I-Application
applications	O
include	O
voice	B-Application
user	I-Application
interfaces	I-Application
such	O
as	O
voice	B-Application
dialing	I-Application
(	O
e.g.	O
</s>
<s>
a	O
radiology	O
report	O
)	O
,	O
determining	O
speaker	O
characteristics	O
,	O
speech-to-text	B-Application
processing	O
(	O
e.g.	O
,	O
word	B-General_Concept
processors	I-General_Concept
or	O
emails	O
)	O
,	O
and	O
aircraft	O
(	O
usually	O
termed	O
direct	B-General_Concept
voice	I-General_Concept
input	I-General_Concept
)	O
.	O
</s>
<s>
The	O
term	O
voice	O
recognition	O
or	O
speaker	B-Application
identification	I-Application
refers	O
to	O
identifying	O
the	O
speaker	O
,	O
rather	O
than	O
what	O
they	O
are	O
saying	O
.	O
</s>
<s>
Recognizing	B-Application
the	I-Application
speaker	I-Application
can	O
simplify	O
the	O
task	O
of	O
translating	O
speech	O
in	O
systems	O
that	O
have	O
been	O
trained	O
on	O
a	O
specific	O
person	O
's	O
voice	O
or	O
it	O
can	O
be	O
used	O
to	O
authenticate	O
or	O
verify	O
the	O
identity	O
of	O
a	O
speaker	O
as	O
part	O
of	O
a	O
security	O
process	O
.	O
</s>
<s>
From	O
the	O
technology	O
perspective	O
,	O
speech	B-Application
recognition	I-Application
has	O
a	O
long	O
history	O
with	O
several	O
waves	O
of	O
major	O
innovations	O
.	O
</s>
<s>
Most	O
recently	O
,	O
the	O
field	O
has	O
benefited	O
from	O
advances	O
in	O
deep	B-Algorithm
learning	I-Algorithm
and	O
big	B-Application
data	I-Application
.	O
</s>
<s>
The	O
advances	O
are	O
evidenced	O
not	O
only	O
by	O
the	O
surge	O
of	O
academic	O
papers	O
published	O
in	O
the	O
field	O
,	O
but	O
more	O
importantly	O
by	O
the	O
worldwide	O
industry	O
adoption	O
of	O
a	O
variety	O
of	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
in	O
designing	O
and	O
deploying	O
speech	B-Application
recognition	I-Application
systems	O
.	O
</s>
<s>
1960	O
–	O
Gunnar	O
Fant	O
developed	O
and	O
published	O
the	O
source-filter	B-Application
model	I-Application
of	I-Application
speech	I-Application
production	I-Application
.	O
</s>
<s>
1962	O
–	O
IBM	O
demonstrated	O
its	O
16-word	O
"	O
Shoebox	O
"	O
machine	O
's	O
speech	B-Application
recognition	I-Application
capability	O
at	O
the	O
1962	O
World	O
's	O
Fair	O
.	O
</s>
<s>
1966	O
–	O
Linear	B-Algorithm
predictive	I-Algorithm
coding	I-Algorithm
(	O
LPC	O
)	O
,	O
a	O
speech	B-Application
coding	I-Application
method	O
,	O
was	O
first	O
proposed	O
by	O
Fumitada	O
Itakura	O
of	O
Nagoya	O
University	O
and	O
Shuzo	O
Saito	O
of	O
Nippon	O
Telegraph	O
and	O
Telephone	O
(	O
NTT	O
)	O
,	O
while	O
working	O
on	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
1969	O
–	O
Funding	O
at	O
Bell	O
Labs	O
dried	O
up	O
for	O
several	O
years	O
when	O
,	O
in	O
1969	O
,	O
the	O
influential	O
John	O
Pierce	O
wrote	O
an	O
open	O
letter	O
that	O
was	O
critical	O
of	O
and	O
defunded	O
speech	B-Application
recognition	I-Application
research	O
.	O
</s>
<s>
Raj	O
Reddy	O
was	O
the	O
first	O
person	O
to	O
take	O
on	O
continuous	O
speech	B-Application
recognition	I-Application
as	O
a	O
graduate	O
student	O
at	O
Stanford	O
University	O
in	O
the	O
late	O
1960s	O
.	O
</s>
<s>
Reddy	O
's	O
system	O
issued	O
spoken	O
commands	O
for	O
playing	O
chess	B-Application
.	O
</s>
<s>
Around	O
this	O
time	O
Soviet	O
researchers	O
invented	O
the	O
dynamic	B-Algorithm
time	I-Algorithm
warping	I-Algorithm
(	O
DTW	O
)	O
algorithm	O
and	O
used	O
it	O
to	O
create	O
a	O
recognizer	O
capable	O
of	O
operating	O
on	O
a	O
200-word	O
vocabulary	O
.	O
</s>
<s>
1971	O
–	O
DARPA	O
funded	O
five	O
years	O
for	O
Speech	B-Application
Understanding	I-Application
Research	O
,	O
speech	B-Application
recognition	I-Application
research	O
seeking	O
a	O
minimum	O
vocabulary	O
size	O
of	O
1,000	O
words	O
.	O
</s>
<s>
They	O
thought	O
speech	B-Application
understanding	I-Application
would	O
be	O
key	O
to	O
making	O
progress	O
in	O
speech	B-Application
recognition	I-Application
,	O
but	O
this	O
later	O
proved	O
untrue	O
.	O
</s>
<s>
This	O
revived	O
speech	B-Application
recognition	I-Application
research	O
post	O
John	O
Pierce	O
's	O
letter	O
.	O
</s>
<s>
1976	O
–	O
The	O
first	O
ICASSP	O
was	O
held	O
in	O
Philadelphia	O
,	O
which	O
since	O
then	O
has	O
been	O
a	O
major	O
venue	O
for	O
the	O
publication	O
of	O
research	O
on	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
A	O
decade	O
later	O
,	O
at	O
CMU	O
,	O
Raj	O
Reddy	O
's	O
students	O
James	O
Baker	O
and	O
Janet	O
M	O
.	O
Baker	O
began	O
using	O
the	O
Hidden	O
Markov	O
Model	O
(	O
HMM	O
)	O
for	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
By	O
the	O
mid-1980s	O
IBM	O
's	O
Fred	O
Jelinek	O
's	O
team	O
created	O
a	O
voice	B-Application
activated	I-Application
typewriter	O
called	O
Tangora	O
,	O
which	O
could	O
handle	O
a	O
20,000	O
-word	O
vocabulary	O
Jelinek	O
's	O
statistical	O
approach	O
put	O
less	O
emphasis	O
on	O
emulating	O
the	O
way	O
the	O
human	O
brain	O
processes	O
and	O
understands	O
speech	O
in	O
favor	O
of	O
using	O
statistical	O
modeling	O
techniques	O
like	O
HMMs	O
.	O
</s>
<s>
However	O
,	O
the	O
HMM	O
proved	O
to	O
be	O
a	O
highly	O
useful	O
way	O
for	O
modeling	O
speech	O
and	O
replaced	O
dynamic	B-Algorithm
time	I-Algorithm
warping	I-Algorithm
to	O
become	O
the	O
dominant	O
speech	B-Application
recognition	I-Application
algorithm	O
in	O
the	O
1980s	O
.	O
</s>
<s>
1982	O
–	O
Dragon	B-Application
Systems	I-Application
,	O
founded	O
by	O
James	O
and	O
Janet	O
M	O
.	O
Baker	O
,	O
was	O
one	O
of	O
IBM	O
's	O
few	O
competitors	O
.	O
</s>
<s>
The	O
1980s	O
also	O
saw	O
the	O
introduction	O
of	O
the	O
n-gram	B-Language
language	I-Language
model	I-Language
.	O
</s>
<s>
1987	O
–	O
The	O
back-off	B-General_Concept
model	I-General_Concept
allowed	O
language	B-Language
models	I-Language
to	O
use	O
multiple	O
length	O
n-grams	B-Language
,	O
and	O
CSELT	O
used	O
HMM	O
to	O
recognize	O
languages	O
(	O
both	O
in	O
software	O
and	O
in	O
hardware	O
specialized	O
processors	O
,	O
e.g.	O
</s>
<s>
RIPAC	B-General_Concept
)	O
.	O
</s>
<s>
At	O
the	O
end	O
of	O
the	O
DARPA	O
program	O
in	O
1976	O
,	O
the	O
best	O
computer	O
available	O
to	O
researchers	O
was	O
the	O
PDP-10	B-Device
with	O
4	O
MB	O
ram	B-Architecture
.	O
</s>
<s>
1984	O
–	O
was	O
released	O
the	O
Apricot	B-Device
Portable	I-Device
with	O
up	O
to	O
4096	O
words	O
support	O
,	O
of	O
which	O
only	O
64	O
could	O
be	O
held	O
in	O
RAM	B-Architecture
at	O
a	O
time	O
.	O
</s>
<s>
By	O
this	O
point	O
,	O
the	O
vocabulary	O
of	O
the	O
typical	O
commercial	O
speech	B-Application
recognition	I-Application
system	O
was	O
larger	O
than	O
the	O
average	O
human	O
vocabulary	O
.	O
</s>
<s>
Raj	O
Reddy	O
's	O
former	O
student	O
,	O
Xuedong	O
Huang	O
,	O
developed	O
the	O
Sphinx-II	B-General_Concept
system	O
at	O
CMU	O
.	O
</s>
<s>
The	O
Sphinx-II	B-General_Concept
system	O
was	O
the	O
first	O
to	O
do	O
speaker-independent	O
,	O
large	O
vocabulary	O
,	O
continuous	O
speech	B-Application
recognition	I-Application
and	O
it	O
had	O
the	O
best	O
performance	O
in	O
DARPA	O
's	O
1992	O
evaluation	O
.	O
</s>
<s>
Handling	O
continuous	O
speech	O
with	O
a	O
large	O
vocabulary	O
was	O
a	O
major	O
milestone	O
in	O
the	O
history	O
of	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
Huang	O
went	O
on	O
to	O
found	O
the	O
speech	B-General_Concept
recognition	I-General_Concept
group	I-General_Concept
at	I-General_Concept
Microsoft	I-General_Concept
in	O
1993	O
.	O
</s>
<s>
Lernout	O
&	O
Hauspie	O
,	O
a	O
Belgium-based	O
speech	B-Application
recognition	I-Application
company	O
,	O
acquired	O
several	O
other	O
companies	O
,	O
including	O
Kurzweil	O
Applied	O
Intelligence	O
in	O
1997	O
and	O
Dragon	B-Application
Systems	I-Application
in	O
2000	O
.	O
</s>
<s>
The	O
L&H	O
speech	B-Application
technology	I-Application
was	O
used	O
in	O
the	O
Windows	B-Application
XP	I-Application
operating	I-Application
system	I-Application
.	O
</s>
<s>
The	O
speech	B-Application
technology	I-Application
from	O
L&H	O
was	O
bought	O
by	O
ScanSoft	O
which	O
became	O
Nuance	O
in	O
2005	O
.	O
</s>
<s>
Apple	O
originally	O
licensed	O
software	O
from	O
Nuance	O
to	O
provide	O
speech	B-Application
recognition	I-Application
capability	O
to	O
its	O
digital	O
assistant	O
Siri	B-Application
.	O
</s>
<s>
In	O
the	O
2000s	O
DARPA	O
sponsored	O
two	O
speech	B-Application
recognition	I-Application
programs	O
:	O
Effective	O
Affordable	O
Reusable	O
Speech-to-Text	B-Application
(	O
EARS	O
)	O
in	O
2002	O
and	O
Global	O
Autonomous	O
Language	O
Exploitation	O
(	O
GALE	O
)	O
.	O
</s>
<s>
Google	B-Application
's	I-Application
first	O
effort	O
at	O
speech	B-Application
recognition	I-Application
came	O
in	O
2007	O
after	O
hiring	O
some	O
researchers	O
from	O
Nuance	O
.	O
</s>
<s>
The	O
recordings	O
from	O
GOOG-411	O
produced	O
valuable	O
data	O
that	O
helped	O
Google	B-Application
improve	O
their	O
recognition	O
systems	O
.	O
</s>
<s>
Google	B-Application
Voice	I-Application
Search	I-Application
is	O
now	O
supported	O
in	O
over	O
30	O
languages	O
.	O
</s>
<s>
In	O
the	O
United	O
States	O
,	O
the	O
National	O
Security	O
Agency	O
has	O
made	O
use	O
of	O
a	O
type	O
of	O
speech	B-Application
recognition	I-Application
for	O
keyword	O
spotting	O
since	O
at	O
least	O
2006	O
.	O
</s>
<s>
Some	O
government	O
research	O
programs	O
focused	O
on	O
intelligence	O
applications	O
of	O
speech	B-Application
recognition	I-Application
,	O
e.g.	O
</s>
<s>
In	O
the	O
early	O
2000s	O
,	O
speech	B-Application
recognition	I-Application
was	O
still	O
dominated	O
by	O
traditional	O
approaches	O
such	O
as	O
Hidden	O
Markov	O
Models	O
combined	O
with	O
feedforward	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
.	O
</s>
<s>
Today	O
,	O
however	O
,	O
many	O
aspects	O
of	O
speech	B-Application
recognition	I-Application
have	O
been	O
taken	O
over	O
by	O
a	O
deep	B-Algorithm
learning	I-Algorithm
method	O
called	O
Long	B-Algorithm
short-term	I-Algorithm
memory	I-Algorithm
(	O
LSTM	B-Algorithm
)	O
,	O
a	O
recurrent	B-Algorithm
neural	I-Algorithm
network	I-Algorithm
published	O
by	O
Sepp	O
Hochreiter	O
&	O
Jürgen	O
Schmidhuber	O
in	O
1997	O
.	O
</s>
<s>
LSTM	B-Algorithm
RNNs	O
avoid	O
the	O
vanishing	B-Algorithm
gradient	I-Algorithm
problem	I-Algorithm
and	O
can	O
learn	O
"	O
Very	O
Deep	B-Algorithm
Learning	I-Algorithm
"	O
tasks	O
that	O
require	O
memories	O
of	O
events	O
that	O
happened	O
thousands	O
of	O
discrete	O
time	O
steps	O
ago	O
,	O
which	O
is	O
important	O
for	O
speech	O
.	O
</s>
<s>
Around	O
2007	O
,	O
LSTM	B-Algorithm
trained	O
by	O
Connectionist	B-Algorithm
Temporal	I-Algorithm
Classification	I-Algorithm
(	O
CTC	O
)	O
started	O
to	O
outperform	O
traditional	O
speech	B-Application
recognition	I-Application
in	O
certain	O
applications	O
.	O
</s>
<s>
In	O
2015	O
,	O
Google	B-Application
's	I-Application
speech	B-Application
recognition	I-Application
reportedly	O
experienced	O
a	O
dramatic	O
performance	O
jump	O
of	O
49%	O
through	O
CTC-trained	O
LSTM	B-Algorithm
,	O
which	O
is	O
now	O
available	O
through	O
Google	B-Application
Voice	I-Application
to	O
all	O
smartphone	B-Application
users	O
.	O
</s>
<s>
Transformers	B-Algorithm
,	O
a	O
type	O
of	O
neural	B-Architecture
network	I-Architecture
based	O
on	O
solely	O
on	O
attention	O
,	O
have	O
been	O
widely	O
adopted	O
in	O
computer	O
vision	O
and	O
language	B-Language
modeling	I-Language
,	O
sparking	O
the	O
interest	O
of	O
adapting	O
such	O
models	O
to	O
new	O
domains	O
,	O
including	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
Some	O
recent	O
papers	O
reported	O
superior	O
performance	O
levels	O
using	O
transformer	B-Algorithm
models	I-Algorithm
for	O
speech	B-Application
recognition	I-Application
,	O
but	O
these	O
models	O
usually	O
require	O
large	O
scale	O
training	O
datasets	O
to	O
reach	O
high	O
performance	O
levels	O
.	O
</s>
<s>
The	O
use	O
of	O
deep	O
feedforward	O
(	O
non-recurrent	O
)	O
networks	O
for	O
acoustic	O
modeling	O
was	O
introduced	O
during	O
the	O
later	O
part	O
of	O
2009	O
by	O
Geoffrey	O
Hinton	O
and	O
his	O
students	O
at	O
the	O
University	O
of	O
Toronto	O
and	O
by	O
Li	O
Deng	O
and	O
colleagues	O
at	O
Microsoft	O
Research	O
,	O
initially	O
in	O
the	O
collaborative	O
work	O
between	O
Microsoft	O
and	O
the	O
University	O
of	O
Toronto	O
which	O
was	O
subsequently	O
expanded	O
to	O
include	O
IBM	O
and	O
Google	B-Application
(	O
hence	O
"	O
The	O
shared	O
views	O
of	O
four	O
research	O
groups	O
"	O
subtitle	O
in	O
their	O
2012	O
review	O
paper	O
)	O
.	O
</s>
<s>
In	O
contrast	O
to	O
the	O
steady	O
incremental	O
improvements	O
of	O
the	O
past	O
few	O
decades	O
,	O
the	O
application	O
of	O
deep	B-Algorithm
learning	I-Algorithm
decreased	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
by	O
30%	O
.	O
</s>
<s>
Researchers	O
have	O
begun	O
to	O
use	O
deep	B-Algorithm
learning	I-Algorithm
techniques	O
for	O
language	B-Language
modeling	I-Language
as	O
well	O
.	O
</s>
<s>
In	O
the	O
long	O
history	O
of	O
speech	B-Application
recognition	I-Application
,	O
both	O
shallow	O
form	O
and	O
deep	O
form	O
(	O
e.g.	O
</s>
<s>
recurrent	O
nets	O
)	O
of	O
artificial	B-Architecture
neural	I-Architecture
networks	I-Architecture
had	O
been	O
explored	O
for	O
many	O
years	O
during	O
1980s	O
,	O
1990s	O
and	O
a	O
few	O
years	O
into	O
the	O
2000s	O
.	O
</s>
<s>
Most	O
speech	B-Application
recognition	I-Application
researchers	O
who	O
understood	O
such	O
barriers	O
hence	O
subsequently	O
moved	O
away	O
from	O
neural	B-Architecture
nets	I-Architecture
to	O
pursue	O
generative	O
modeling	O
approaches	O
until	O
the	O
recent	O
resurgence	O
of	O
deep	B-Algorithm
learning	I-Algorithm
starting	O
around	O
2009	O
–	O
2010	O
that	O
had	O
overcome	O
all	O
these	O
difficulties	O
.	O
</s>
<s>
reviewed	O
part	O
of	O
this	O
recent	O
history	O
about	O
how	O
their	O
collaboration	O
with	O
each	O
other	O
and	O
then	O
with	O
colleagues	O
across	O
four	O
groups	O
(	O
University	O
of	O
Toronto	O
,	O
Microsoft	O
,	O
Google	B-Application
,	O
and	O
IBM	O
)	O
ignited	O
a	O
renaissance	O
of	O
applications	O
of	O
deep	O
feedforward	O
neural	B-Architecture
networks	I-Architecture
to	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
By	O
early	O
2010s	O
speech	B-Application
recognition	I-Application
,	O
also	O
called	O
voice	O
recognition	O
was	O
clearly	O
differentiated	O
from	O
speaker	B-Application
recognition	I-Application
,	O
and	O
speaker	O
independence	O
was	O
considered	O
a	O
major	O
breakthrough	O
.	O
</s>
<s>
Multiple	O
deep	B-Algorithm
learning	I-Algorithm
models	O
were	O
used	O
to	O
optimize	O
speech	B-Application
recognition	I-Application
accuracy	O
.	O
</s>
<s>
The	O
speech	B-Application
recognition	I-Application
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
was	O
reported	O
to	O
be	O
as	O
low	O
as	O
4	O
professional	O
human	O
transcribers	O
working	O
together	O
on	O
the	O
same	O
benchmark	O
,	O
which	O
was	O
funded	O
by	O
IBM	O
Watson	O
speech	O
team	O
on	O
the	O
same	O
task	O
.	O
</s>
<s>
Both	O
acoustic	O
modeling	O
and	O
language	B-Language
modeling	I-Language
are	O
important	O
parts	O
of	O
modern	O
statistically	O
based	O
speech	B-Application
recognition	I-Application
algorithms	O
.	O
</s>
<s>
Language	B-Language
modeling	I-Language
is	O
also	O
used	O
in	O
many	O
other	O
natural	B-Language
language	I-Language
processing	I-Language
applications	O
such	O
as	O
document	B-Algorithm
classification	I-Algorithm
or	O
statistical	B-General_Concept
machine	I-General_Concept
translation	I-General_Concept
.	O
</s>
<s>
Modern	O
general-purpose	O
speech	B-Application
recognition	I-Application
systems	O
are	O
based	O
on	O
hidden	O
Markov	O
models	O
.	O
</s>
<s>
HMMs	O
are	O
used	O
in	O
speech	B-Application
recognition	I-Application
because	O
a	O
speech	O
signal	O
can	O
be	O
viewed	O
as	O
a	O
piecewise	O
stationary	O
signal	O
or	O
a	O
short-time	O
stationary	O
signal	O
.	O
</s>
<s>
In	O
a	O
short	O
time	O
scale	O
(	O
e.g.	O
,	O
10	O
milliseconds	O
)	O
,	O
speech	O
can	O
be	O
approximated	O
as	O
a	O
stationary	B-Algorithm
process	I-Algorithm
.	O
</s>
<s>
In	O
speech	B-Application
recognition	I-Application
,	O
the	O
hidden	O
Markov	O
model	O
would	O
output	O
a	O
sequence	O
of	O
n-dimensional	O
real-valued	O
vectors	O
(	O
with	O
n	O
being	O
a	O
small	O
integer	O
,	O
such	O
as	O
10	O
)	O
,	O
outputting	O
one	O
of	O
these	O
every	O
10	O
milliseconds	O
.	O
</s>
<s>
The	O
vectors	O
would	O
consist	O
of	O
cepstral	B-Algorithm
coefficients	O
,	O
which	O
are	O
obtained	O
by	O
taking	O
a	O
Fourier	B-Algorithm
transform	I-Algorithm
of	O
a	O
short	O
time	O
window	O
of	O
speech	O
and	O
decorrelating	O
the	O
spectrum	O
using	O
a	O
cosine	O
transform	O
,	O
then	O
taking	O
the	O
first	O
(	O
most	O
significant	O
)	O
coefficients	O
.	O
</s>
<s>
Each	O
word	O
,	O
or	O
(	O
for	O
more	O
general	O
speech	B-Application
recognition	I-Application
systems	O
)	O
,	O
each	O
phoneme	B-Language
,	O
will	O
have	O
a	O
different	O
output	O
distribution	O
;	O
a	O
hidden	O
Markov	O
model	O
for	O
a	O
sequence	O
of	O
words	O
or	O
phonemes	B-Language
is	O
made	O
by	O
concatenating	O
the	O
individual	O
trained	O
hidden	O
Markov	O
models	O
for	O
the	O
separate	O
words	O
and	O
phonemes	B-Language
.	O
</s>
<s>
Described	O
above	O
are	O
the	O
core	O
elements	O
of	O
the	O
most	O
common	O
,	O
HMM-based	O
approach	O
to	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
Modern	O
speech	B-Application
recognition	I-Application
systems	O
use	O
various	O
combinations	O
of	O
a	O
number	O
of	O
standard	O
techniques	O
in	O
order	O
to	O
improve	O
results	O
over	O
the	O
basic	O
approach	O
described	O
above	O
.	O
</s>
<s>
A	O
typical	O
large-vocabulary	O
system	O
would	O
need	O
context	O
dependency	O
for	O
the	O
phonemes	B-Language
(	O
so	O
phonemes	B-Language
with	O
different	O
left	O
and	O
right	O
context	O
have	O
different	O
realizations	O
as	O
HMM	O
states	O
)	O
;	O
it	O
would	O
use	O
cepstral	B-Algorithm
normalization	O
to	O
normalize	O
for	O
a	O
different	O
speaker	O
and	O
recording	O
conditions	O
;	O
for	O
further	O
speaker	O
normalization	O
,	O
it	O
might	O
use	O
vocal	O
tract	O
length	O
normalization	O
(	O
VTLN	O
)	O
for	O
male-female	O
normalization	O
and	O
maximum	O
likelihood	O
linear	O
regression	O
(	O
MLLR	O
)	O
for	O
more	O
general	O
speaker	O
adaptation	O
.	O
</s>
<s>
The	O
features	O
would	O
have	O
so-called	O
delta	O
and	O
delta-delta	O
coefficients	O
to	O
capture	O
speech	O
dynamics	O
and	O
in	O
addition	O
,	O
might	O
use	O
heteroscedastic	B-General_Concept
linear	B-General_Concept
discriminant	I-General_Concept
analysis	I-General_Concept
(	O
HLDA	O
)	O
;	O
or	O
might	O
skip	O
the	O
delta	O
and	O
delta-delta	O
coefficients	O
and	O
use	O
splicing	O
and	O
an	O
LDA-based	O
projection	O
followed	O
perhaps	O
by	O
heteroscedastic	B-General_Concept
linear	B-General_Concept
discriminant	I-General_Concept
analysis	I-General_Concept
or	O
a	O
global	O
semi-tied	O
co	O
variance	O
transform	O
(	O
also	O
known	O
as	O
maximum	O
likelihood	O
linear	O
transform	O
,	O
or	O
MLLT	O
)	O
.	O
</s>
<s>
Decoding	O
of	O
the	O
speech	O
(	O
the	O
term	O
for	O
what	O
happens	O
when	O
the	O
system	O
is	O
presented	O
with	O
a	O
new	O
utterance	O
and	O
must	O
compute	O
the	O
most	O
likely	O
source	O
sentence	O
)	O
would	O
probably	O
use	O
the	O
Viterbi	B-Algorithm
algorithm	I-Algorithm
to	O
find	O
the	O
best	O
path	O
,	O
and	O
here	O
there	O
is	O
a	O
choice	O
between	O
dynamically	O
creating	O
a	O
combination	O
hidden	O
Markov	O
model	O
,	O
which	O
includes	O
both	O
the	O
acoustic	O
and	O
language	B-Language
model	I-Language
information	O
and	O
combining	O
it	O
statically	O
beforehand	O
(	O
the	O
finite	B-Architecture
state	I-Architecture
transducer	I-Architecture
,	O
or	O
FST	O
,	O
approach	O
)	O
.	O
</s>
<s>
Re	O
scoring	O
is	O
usually	O
done	O
by	O
trying	O
to	O
minimize	O
the	O
Bayes	B-General_Concept
risk	I-General_Concept
(	O
or	O
an	O
approximation	O
thereof	O
)	O
:	O
Instead	O
of	O
taking	O
the	O
source	O
sentence	O
with	O
maximal	O
probability	O
,	O
we	O
try	O
to	O
take	O
the	O
sentence	O
that	O
minimizes	O
the	O
expectancy	O
of	O
a	O
given	O
loss	O
function	O
with	O
regards	O
to	O
all	O
possible	O
transcriptions	O
(	O
i.e.	O
,	O
we	O
take	O
the	O
sentence	O
that	O
minimizes	O
the	O
average	O
distance	O
to	O
other	O
possible	O
sentences	O
weighted	O
by	O
their	O
estimated	O
probability	O
)	O
.	O
</s>
<s>
Efficient	O
algorithms	O
have	O
been	O
devised	O
to	O
re	O
score	O
lattices	O
represented	O
as	O
weighted	O
finite	B-Architecture
state	I-Architecture
transducers	I-Architecture
with	O
edit	O
distances	O
represented	O
themselves	O
as	O
a	O
finite	B-Architecture
state	I-Architecture
transducer	I-Architecture
verifying	O
certain	O
assumptions	O
.	O
</s>
<s>
Dynamic	B-Algorithm
time	I-Algorithm
warping	I-Algorithm
is	O
an	O
approach	O
that	O
was	O
historically	O
used	O
for	O
speech	B-Application
recognition	I-Application
but	O
has	O
now	O
largely	O
been	O
displaced	O
by	O
the	O
more	O
successful	O
HMM-based	O
approach	O
.	O
</s>
<s>
Dynamic	B-Algorithm
time	I-Algorithm
warping	I-Algorithm
is	O
an	O
algorithm	O
for	O
measuring	O
similarity	O
between	O
two	O
sequences	O
that	O
may	O
vary	O
in	O
time	O
or	O
speed	O
.	O
</s>
<s>
A	O
well-known	O
application	O
has	O
been	O
automatic	B-Application
speech	I-Application
recognition	I-Application
,	O
to	O
cope	O
with	O
different	O
speaking	O
speeds	O
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
emerged	O
as	O
an	O
attractive	O
acoustic	O
modeling	O
approach	O
in	O
ASR	O
in	O
the	O
late	O
1980s	O
.	O
</s>
<s>
Since	O
then	O
,	O
neural	B-Architecture
networks	I-Architecture
have	O
been	O
used	O
in	O
many	O
aspects	O
of	O
speech	B-Application
recognition	I-Application
such	O
as	O
phoneme	B-Language
classification	O
,	O
phoneme	B-Language
classification	O
through	O
multi-objective	O
evolutionary	O
algorithms	O
,	O
isolated	O
word	O
recognition	O
,	O
audiovisual	B-General_Concept
speech	I-General_Concept
recognition	I-General_Concept
,	O
audiovisual	O
speaker	B-Application
recognition	I-Application
and	O
speaker	O
adaptation	O
.	O
</s>
<s>
Neural	B-Architecture
networks	I-Architecture
make	O
fewer	O
explicit	O
assumptions	O
about	O
feature	O
statistical	O
properties	O
than	O
HMMs	O
and	O
have	O
several	O
qualities	O
making	O
them	O
attractive	O
recognition	O
models	O
for	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
When	O
used	O
to	O
estimate	O
the	O
probabilities	O
of	O
a	O
speech	O
feature	O
segment	O
,	O
neural	B-Architecture
networks	I-Architecture
allow	O
discriminative	O
training	O
in	O
a	O
natural	O
and	O
efficient	O
manner	O
.	O
</s>
<s>
However	O
,	O
in	O
spite	O
of	O
their	O
effectiveness	O
in	O
classifying	O
short-time	O
units	O
such	O
as	O
individual	O
phonemes	B-Language
and	O
isolated	O
words	O
,	O
early	O
neural	B-Architecture
networks	I-Architecture
were	O
rarely	O
successful	O
for	O
continuous	O
recognition	O
tasks	O
because	O
of	O
their	O
limited	O
ability	O
to	O
model	O
temporal	O
dependencies	O
.	O
</s>
<s>
One	O
approach	O
to	O
this	O
limitation	O
was	O
to	O
use	O
neural	B-Architecture
networks	I-Architecture
as	O
a	O
pre-processing	O
,	O
feature	O
transformation	O
or	O
dimensionality	O
reduction	O
,	O
step	O
prior	O
to	O
HMM	O
based	O
recognition	O
.	O
</s>
<s>
However	O
,	O
more	O
recently	O
,	O
LSTM	B-Algorithm
and	O
related	O
recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
(	O
RNNs	O
)	O
,	O
Time	O
Delay	O
Neural	O
Networks(TDNN's )	O
,	O
and	O
transformers	B-Algorithm
.	O
</s>
<s>
Deep	O
Neural	B-Architecture
Networks	I-Architecture
and	O
Denoising	O
Autoencoders	B-Algorithm
are	O
also	O
under	O
investigation	O
.	O
</s>
<s>
A	O
deep	O
feedforward	O
neural	B-Architecture
network	I-Architecture
(	O
DNN	O
)	O
is	O
an	O
artificial	B-Architecture
neural	I-Architecture
network	I-Architecture
with	O
multiple	O
hidden	O
layers	O
of	O
units	O
between	O
the	O
input	O
and	O
output	O
layers	O
.	O
</s>
<s>
Similar	O
to	O
shallow	O
neural	B-Architecture
networks	I-Architecture
,	O
DNNs	O
can	O
model	O
complex	O
non-linear	O
relationships	O
.	O
</s>
<s>
A	O
success	O
of	O
DNNs	O
in	O
large	O
vocabulary	O
speech	B-Application
recognition	I-Application
occurred	O
in	O
2010	O
by	O
industrial	O
researchers	O
,	O
in	O
collaboration	O
with	O
academic	O
researchers	O
,	O
where	O
large	O
output	O
layers	O
of	O
the	O
DNN	O
based	O
on	O
context	O
dependent	O
HMM	O
states	O
constructed	O
by	O
decision	O
trees	O
were	O
adopted	O
.	O
</s>
<s>
One	O
fundamental	O
principle	O
of	O
deep	B-Algorithm
learning	I-Algorithm
is	O
to	O
do	O
away	O
with	O
hand-crafted	O
feature	B-General_Concept
engineering	I-General_Concept
and	O
to	O
use	O
raw	O
features	O
.	O
</s>
<s>
This	O
principle	O
was	O
first	O
explored	O
successfully	O
in	O
the	O
architecture	O
of	O
deep	O
autoencoder	B-Algorithm
on	O
the	O
"	O
raw	O
"	O
spectrogram	O
or	O
linear	O
filter-bank	O
features	O
,	O
showing	O
its	O
superiority	O
over	O
the	O
Mel-Cepstral	O
features	O
which	O
contain	O
a	O
few	O
stages	O
of	O
fixed	O
transformation	O
from	O
spectrograms	O
.	O
</s>
<s>
The	O
true	O
"	O
raw	O
"	O
features	O
of	O
speech	O
,	O
waveforms	O
,	O
have	O
more	O
recently	O
been	O
shown	O
to	O
produce	O
excellent	O
larger-scale	O
speech	B-Application
recognition	I-Application
results	O
.	O
</s>
<s>
Traditional	O
phonetic-based	O
(	O
i.e.	O
,	O
all	O
HMM-based	O
model	O
)	O
approaches	O
required	O
separate	O
components	O
and	O
training	O
for	O
the	O
pronunciation	O
,	O
acoustic	O
,	O
and	O
language	B-Language
model	I-Language
.	O
</s>
<s>
End-to-end	O
models	O
jointly	O
learn	O
all	O
the	O
components	O
of	O
the	O
speech	B-Application
recognizer	I-Application
.	O
</s>
<s>
For	O
example	O
,	O
a	O
n-gram	B-Language
language	I-Language
model	I-Language
is	O
required	O
for	O
all	O
HMM-based	O
systems	O
,	O
and	O
a	O
typical	O
n-gram	B-Language
language	I-Language
model	I-Language
often	O
takes	O
several	O
gigabytes	O
in	O
memory	O
making	O
them	O
impractical	O
to	O
deploy	O
on	O
mobile	O
devices	O
.	O
</s>
<s>
Consequently	O
,	O
modern	O
commercial	O
ASR	O
systems	O
from	O
Google	B-Application
and	O
Apple	O
(	O
)	O
are	O
deployed	O
on	O
the	O
cloud	O
and	O
require	O
a	O
network	O
connection	O
as	O
opposed	O
to	O
the	O
device	O
locally	O
.	O
</s>
<s>
The	O
first	O
attempt	O
at	O
end-to-end	O
ASR	O
was	O
with	O
Connectionist	B-Algorithm
Temporal	I-Algorithm
Classification	I-Algorithm
(	O
CTC	O
)	O
-based	O
systems	O
introduced	O
by	O
Alex	O
Graves	O
of	O
Google	B-Application
DeepMind	I-Application
and	O
Navdeep	O
Jaitly	O
of	O
the	O
University	O
of	O
Toronto	O
in	O
2014	O
.	O
</s>
<s>
The	O
model	O
consisted	O
of	O
recurrent	B-Algorithm
neural	I-Algorithm
networks	I-Algorithm
and	O
a	O
CTC	O
layer	O
.	O
</s>
<s>
Jointly	O
,	O
the	O
RNN-CTC	O
model	O
learns	O
the	O
pronunciation	O
and	O
acoustic	B-General_Concept
model	I-General_Concept
together	O
,	O
however	O
it	O
is	O
incapable	O
of	O
learning	O
the	O
language	O
due	O
to	O
conditional	O
independence	O
assumptions	O
similar	O
to	O
a	O
HMM	O
.	O
</s>
<s>
Consequently	O
,	O
CTC	O
models	O
can	O
directly	O
learn	O
to	O
map	O
speech	O
acoustics	O
to	O
English	O
characters	O
,	O
but	O
the	O
models	O
make	O
many	O
common	O
spelling	O
mistakes	O
and	O
must	O
rely	O
on	O
a	O
separate	O
language	B-Language
model	I-Language
to	O
clean	O
up	O
the	O
transcripts	O
.	O
</s>
<s>
Later	O
,	O
Baidu	B-Application
expanded	O
on	O
the	O
work	O
with	O
extremely	O
large	O
datasets	O
and	O
demonstrated	O
some	O
commercial	O
success	O
in	O
Chinese	O
Mandarin	O
and	O
English	O
.	O
</s>
<s>
In	O
2016	O
,	O
University	O
of	O
Oxford	O
presented	O
LipNet	B-Application
,	O
the	O
first	O
end-to-end	O
sentence-level	O
lipreading	O
model	O
,	O
using	O
spatiotemporal	O
convolutions	O
coupled	O
with	O
an	O
RNN-CTC	O
architecture	O
,	O
surpassing	O
human-level	O
performance	O
in	O
a	O
restricted	O
grammar	O
dataset	O
.	O
</s>
<s>
A	O
large-scale	O
CNN-RNN-CTC	O
architecture	O
was	O
presented	O
in	O
2018	O
by	O
Google	B-Application
DeepMind	I-Application
achieving	O
6	O
times	O
better	O
performance	O
than	O
human	O
experts	O
.	O
</s>
<s>
of	O
Carnegie	O
Mellon	O
University	O
and	O
Google	B-Application
Brain	I-Application
and	O
Bahdanau	O
et	O
al	O
.	O
</s>
<s>
Unlike	O
CTC-based	O
models	O
,	O
attention-based	O
models	O
do	O
not	O
have	O
conditional-independence	O
assumptions	O
and	O
can	O
learn	O
all	O
the	O
components	O
of	O
a	O
speech	B-Application
recognizer	I-Application
including	O
the	O
pronunciation	O
,	O
acoustic	O
and	O
language	B-Language
model	I-Language
directly	O
.	O
</s>
<s>
This	O
means	O
,	O
during	O
deployment	O
,	O
there	O
is	O
no	O
need	O
to	O
carry	O
around	O
a	O
language	B-Language
model	I-Language
making	O
it	O
very	O
practical	O
for	O
applications	O
with	O
limited	O
memory	O
.	O
</s>
<s>
By	O
the	O
end	O
of	O
2016	O
,	O
the	O
attention-based	O
models	O
have	O
seen	O
considerable	O
success	O
including	O
outperforming	O
the	O
CTC	O
models	O
(	O
with	O
or	O
without	O
an	O
external	O
language	B-Language
model	I-Language
)	O
.	O
</s>
<s>
Latent	O
Sequence	O
Decompositions	O
(	O
LSD	O
)	O
was	O
proposed	O
by	O
Carnegie	O
Mellon	O
University	O
,	O
MIT	O
and	O
Google	B-Application
Brain	I-Application
to	O
directly	O
emit	O
sub-word	O
units	O
which	O
are	O
more	O
natural	O
than	O
English	O
characters	O
;	O
University	O
of	O
Oxford	O
and	O
Google	B-Application
DeepMind	I-Application
extended	O
LAS	O
to	O
"	O
Watch	O
,	O
Listen	O
,	O
Attend	O
and	O
Spell	O
"	O
(	O
WLAS	O
)	O
to	O
handle	O
lip	O
reading	O
surpassing	O
human-level	O
performance	O
.	O
</s>
<s>
Typically	O
a	O
manual	O
control	O
input	O
,	O
for	O
example	O
by	O
means	O
of	O
a	O
finger	O
control	O
on	O
the	O
steering-wheel	O
,	O
enables	O
the	O
speech	B-Application
recognition	I-Application
system	O
and	O
this	O
is	O
signaled	O
to	O
the	O
driver	O
by	O
an	O
audio	O
prompt	O
.	O
</s>
<s>
Simple	O
voice	B-Application
commands	I-Application
may	O
be	O
used	O
to	O
initiate	O
phone	O
calls	O
,	O
select	O
radio	O
stations	O
or	O
play	O
music	O
from	O
a	O
compatible	O
smartphone	B-Application
,	O
MP3	O
player	O
or	O
music-loaded	O
flash	O
drive	O
.	O
</s>
<s>
Some	O
of	O
the	O
most	O
recent	O
car	O
models	O
offer	O
natural-language	O
speech	B-Application
recognition	I-Application
in	O
place	O
of	O
a	O
fixed	O
set	O
of	O
commands	O
,	O
allowing	O
the	O
driver	O
to	O
use	O
full	O
sentences	O
and	O
common	O
phrases	O
.	O
</s>
<s>
In	O
the	O
health	O
care	O
sector	O
,	O
speech	B-Application
recognition	I-Application
can	O
be	O
implemented	O
in	O
front-end	O
or	O
back-end	O
of	O
the	O
medical	O
documentation	O
process	O
.	O
</s>
<s>
Front-end	O
speech	B-Application
recognition	I-Application
is	O
where	O
the	O
provider	O
dictates	O
into	O
a	O
speech-recognition	B-Application
engine	O
,	O
the	O
recognized	O
words	O
are	O
displayed	O
as	O
they	O
are	O
spoken	O
,	O
and	O
the	O
dictator	O
is	O
responsible	O
for	O
editing	O
and	O
signing	O
off	O
on	O
the	O
document	O
.	O
</s>
<s>
Back-end	O
or	O
deferred	O
speech	B-Application
recognition	I-Application
is	O
where	O
the	O
provider	O
dictates	O
into	O
a	O
digital	O
dictation	O
system	O
,	O
the	O
voice	O
is	O
routed	O
through	O
a	O
speech-recognition	B-Application
machine	O
and	O
the	O
recognized	O
draft	O
document	O
is	O
routed	O
along	O
with	O
the	O
original	O
voice	O
file	O
to	O
the	O
editor	O
,	O
where	O
the	O
draft	O
is	O
edited	O
and	O
report	O
finalized	O
.	O
</s>
<s>
Deferred	O
speech	B-Application
recognition	I-Application
is	O
widely	O
used	O
in	O
the	O
industry	O
currently	O
.	O
</s>
<s>
One	O
of	O
the	O
major	O
issues	O
relating	O
to	O
the	O
use	O
of	O
speech	B-Application
recognition	I-Application
in	O
healthcare	O
is	O
that	O
the	O
American	O
Recovery	O
and	O
Reinvestment	O
Act	O
of	O
2009	O
(	O
ARRA	O
)	O
provides	O
for	O
substantial	O
financial	O
benefits	O
to	O
physicians	O
who	O
utilize	O
an	O
EMR	O
according	O
to	O
"	O
Meaningful	O
Use	O
"	O
standards	O
.	O
</s>
<s>
These	O
standards	O
require	O
that	O
a	O
substantial	O
amount	O
of	O
data	O
be	O
maintained	O
by	O
the	O
EMR	O
(	O
now	O
more	O
commonly	O
referred	O
to	O
as	O
an	O
Electronic	B-Application
Health	I-Application
Record	I-Application
or	O
EHR	O
)	O
.	O
</s>
<s>
The	O
use	O
of	O
speech	B-Application
recognition	I-Application
is	O
more	O
naturally	O
suited	O
to	O
the	O
generation	O
of	O
narrative	O
text	O
,	O
as	O
part	O
of	O
a	O
radiology/pathology	O
interpretation	O
,	O
progress	O
note	O
or	O
discharge	O
summary	O
:	O
the	O
ergonomic	O
gains	O
of	O
using	O
speech	B-Application
recognition	I-Application
to	O
enter	O
structured	O
discrete	O
data	O
(	O
e.g.	O
,	O
numeric	O
values	O
or	O
codes	O
from	O
a	O
list	O
or	O
a	O
controlled	O
vocabulary	O
)	O
are	O
relatively	O
minimal	O
for	O
people	O
who	O
are	O
sighted	O
and	O
who	O
can	O
operate	O
a	O
keyboard	O
and	O
mouse	O
.	O
</s>
<s>
A	O
more	O
significant	O
issue	O
is	O
that	O
most	O
EHRs	B-Application
have	O
not	O
been	O
expressly	O
tailored	O
to	O
take	O
advantage	O
of	O
voice-recognition	O
capabilities	O
.	O
</s>
<s>
A	O
large	O
part	O
of	O
the	O
clinician	O
's	O
interaction	O
with	O
the	O
EHR	O
involves	O
navigation	O
through	O
the	O
user	B-Application
interface	I-Application
using	O
menus	O
,	O
and	O
tab/button	O
clicks	O
,	O
and	O
is	O
heavily	O
dependent	O
on	O
keyboard	O
and	O
mouse	O
:	O
voice-based	O
navigation	O
provides	O
only	O
modest	O
ergonomic	O
benefits	O
.	O
</s>
<s>
Prolonged	O
use	O
of	O
speech	B-Application
recognition	I-Application
software	I-Application
in	O
conjunction	O
with	O
word	B-General_Concept
processors	I-General_Concept
has	O
shown	O
benefits	O
to	O
short-term-memory	O
restrengthening	O
in	O
brain	O
AVM	O
patients	O
who	O
have	O
been	O
treated	O
with	O
resection	O
.	O
</s>
<s>
Substantial	O
efforts	O
have	O
been	O
devoted	O
in	O
the	O
last	O
decade	O
to	O
the	O
test	O
and	O
evaluation	O
of	O
speech	B-Application
recognition	I-Application
in	O
fighter	O
aircraft	O
.	O
</s>
<s>
Of	O
particular	O
note	O
have	O
been	O
the	O
US	O
program	O
in	O
speech	B-Application
recognition	I-Application
for	O
the	O
Advanced	O
Fighter	O
Technology	O
Integration	O
(	O
AFTI	O
)	O
/F	O
-16	O
aircraft	O
(	O
F-16	O
VISTA	O
)	O
,	O
the	O
program	O
in	O
France	O
for	O
Mirage	O
aircraft	O
,	O
and	O
other	O
programs	O
in	O
the	O
UK	O
dealing	O
with	O
a	O
variety	O
of	O
aircraft	O
platforms	O
.	O
</s>
<s>
In	O
these	O
programs	O
,	O
speech	B-Application
recognizers	I-Application
have	O
been	O
operated	O
successfully	O
in	O
fighter	O
aircraft	O
,	O
with	O
applications	O
including	O
setting	O
radio	O
frequencies	O
,	O
commanding	O
an	O
autopilot	O
system	O
,	O
setting	O
steer-point	O
coordinates	O
and	O
weapons	O
release	O
parameters	O
,	O
and	O
controlling	O
flight	O
display	O
.	O
</s>
<s>
Working	O
with	O
Swedish	O
pilots	O
flying	O
in	O
the	O
JAS-39	O
Gripen	O
cockpit	O
,	O
Englund	O
(	O
2004	O
)	O
found	O
recognition	O
deteriorated	O
with	O
increasing	O
g-loads	B-Application
.	O
</s>
<s>
The	O
Eurofighter	B-Application
Typhoon	I-Application
,	O
currently	O
in	O
service	O
with	O
the	O
UK	O
RAF	O
,	O
employs	O
a	O
speaker-dependent	O
system	O
,	O
requiring	O
each	O
pilot	O
to	O
create	O
a	O
template	O
.	O
</s>
<s>
Voice	B-Application
commands	I-Application
are	O
confirmed	O
by	O
visual	O
and/or	O
aural	O
feedback	O
.	O
</s>
<s>
The	O
system	O
is	O
seen	O
as	O
a	O
major	O
design	O
feature	O
in	O
the	O
reduction	O
of	O
pilot	O
workload	O
,	O
and	O
even	O
allows	O
the	O
pilot	O
to	O
assign	O
targets	O
to	O
his	O
aircraft	O
with	O
two	O
simple	O
voice	B-Application
commands	I-Application
or	O
to	O
any	O
of	O
his	O
wingmen	O
with	O
only	O
five	O
commands	O
.	O
</s>
<s>
The	O
problems	O
of	O
achieving	O
high	O
recognition	O
accuracy	O
under	O
stress	O
and	O
noise	O
are	O
particularly	O
relevant	O
in	O
the	O
helicopter	B-Application
environment	O
as	O
well	O
as	O
in	O
the	O
jet	O
fighter	O
environment	O
.	O
</s>
<s>
The	O
acoustic	O
noise	O
problem	O
is	O
actually	O
more	O
severe	O
in	O
the	O
helicopter	B-Application
environment	O
,	O
not	O
only	O
because	O
of	O
the	O
high	O
noise	O
levels	O
but	O
also	O
because	O
the	O
helicopter	B-Application
pilot	O
,	O
in	O
general	O
,	O
does	O
not	O
wear	O
a	O
facemask	O
,	O
which	O
would	O
reduce	O
acoustic	O
noise	O
in	O
the	O
microphone	B-Application
.	O
</s>
<s>
Substantial	O
test	O
and	O
evaluation	O
programs	O
have	O
been	O
carried	O
out	O
in	O
the	O
past	O
decade	O
in	O
speech	B-Application
recognition	I-Application
systems	O
applications	O
in	O
helicopters	B-Application
,	O
notably	O
by	O
the	O
U.S.	O
Army	O
Avionics	O
Research	O
and	O
Development	O
Activity	O
(	O
AVRADA	O
)	O
and	O
by	O
the	O
Royal	O
Aerospace	O
Establishment	O
(	O
RAE	O
)	O
in	O
the	O
UK	O
.	O
</s>
<s>
Work	O
in	O
France	O
has	O
included	O
speech	B-Application
recognition	I-Application
in	O
the	O
Puma	O
helicopter	B-Application
.	O
</s>
<s>
As	O
in	O
fighter	O
applications	O
,	O
the	O
overriding	O
issue	O
for	O
voice	O
in	O
helicopters	B-Application
is	O
the	O
impact	O
on	O
pilot	O
effectiveness	O
.	O
</s>
<s>
Much	O
remains	O
to	O
be	O
done	O
both	O
in	O
speech	B-Application
recognition	I-Application
and	O
in	O
overall	O
speech	B-Application
technology	I-Application
in	O
order	O
to	O
consistently	O
achieve	O
performance	O
improvements	O
in	O
operational	O
settings	O
.	O
</s>
<s>
Training	O
for	O
air	O
traffic	O
controllers	O
(	O
ATC	O
)	O
represents	O
an	O
excellent	O
application	O
for	O
speech	B-Application
recognition	I-Application
systems	O
.	O
</s>
<s>
Speech	B-Application
recognition	I-Application
and	O
synthesis	B-Application
techniques	O
offer	O
the	O
potential	O
to	O
eliminate	O
the	O
need	O
for	O
a	O
person	O
to	O
act	O
as	O
a	O
pseudo-pilot	O
,	O
thus	O
reducing	O
training	O
and	O
support	O
personnel	O
.	O
</s>
<s>
In	O
theory	O
,	O
Air	O
controller	O
tasks	O
are	O
also	O
characterized	O
by	O
highly	O
structured	O
speech	O
as	O
the	O
primary	O
output	O
of	O
the	O
controller	O
,	O
hence	O
reducing	O
the	O
difficulty	O
of	O
the	O
speech	B-Application
recognition	I-Application
task	O
should	O
be	O
possible	O
.	O
</s>
<s>
While	O
this	O
document	O
gives	O
less	O
than	O
150	O
examples	O
of	O
such	O
phrases	O
,	O
the	O
number	O
of	O
phrases	O
supported	O
by	O
one	O
of	O
the	O
simulation	O
vendors	O
speech	B-Application
recognition	I-Application
systems	O
is	O
in	O
excess	O
of	O
500,000	O
.	O
</s>
<s>
The	O
USAF	O
,	O
USMC	O
,	O
US	O
Army	O
,	O
US	O
Navy	O
,	O
and	O
FAA	O
as	O
well	O
as	O
a	O
number	O
of	O
international	O
ATC	O
training	O
organizations	O
such	O
as	O
the	O
Royal	O
Australian	O
Air	O
Force	O
and	O
Civil	O
Aviation	O
Authorities	O
in	O
Italy	O
,	O
Brazil	O
,	O
and	O
Canada	O
are	O
currently	O
using	O
ATC	O
simulators	O
with	O
speech	B-Application
recognition	I-Application
from	O
a	O
number	O
of	O
different	O
vendors	O
.	O
</s>
<s>
The	O
improvement	O
of	O
mobile	O
processor	O
speeds	O
has	O
made	O
speech	B-Application
recognition	I-Application
practical	O
in	O
smartphones	B-Application
.	O
</s>
<s>
Speech	O
is	O
used	O
mostly	O
as	O
a	O
part	O
of	O
a	O
user	B-Application
interface	I-Application
,	O
for	O
creating	O
predefined	O
or	O
custom	O
speech	O
commands	O
.	O
</s>
<s>
For	O
language	O
learning	O
,	O
speech	B-Application
recognition	I-Application
can	O
be	O
useful	O
for	O
learning	O
a	O
second	O
language	O
.	O
</s>
<s>
Students	O
who	O
are	O
physically	O
disabled	O
,	O
have	O
a	O
Repetitive	O
strain	O
injury/other	O
injuries	O
to	O
the	O
upper	O
extremities	O
can	O
be	O
relieved	O
from	O
having	O
to	O
worry	O
about	O
handwriting	O
,	O
typing	O
,	O
or	O
working	O
with	O
scribe	O
on	O
school	O
assignments	O
by	O
using	O
speech-to-text	B-Application
programs	O
.	O
</s>
<s>
They	O
can	O
also	O
utilize	O
speech	B-Application
recognition	I-Application
technology	I-Application
to	O
enjoy	O
searching	O
the	O
Internet	O
or	O
using	O
a	O
computer	O
at	O
home	O
without	O
having	O
to	O
physically	O
operate	O
a	O
mouse	O
and	O
keyboard	O
.	O
</s>
<s>
Speech	B-Application
recognition	I-Application
can	O
allow	O
students	O
with	O
learning	O
disabilities	O
to	O
become	O
better	O
writers	O
.	O
</s>
<s>
The	O
use	O
of	O
voice	B-Application
recognition	I-Application
software	I-Application
,	O
in	O
conjunction	O
with	O
a	O
digital	O
audio	O
recorder	O
and	O
a	O
personal	O
computer	O
running	O
word-processing	B-General_Concept
software	O
has	O
proven	O
to	O
be	O
positive	O
for	O
restoring	O
damaged	O
short-term	O
memory	O
capacity	O
,	O
in	O
stroke	O
and	O
craniotomy	O
individuals	O
.	O
</s>
<s>
People	O
with	O
disabilities	O
can	O
benefit	O
from	O
speech	B-Application
recognition	I-Application
programs	O
.	O
</s>
<s>
For	O
individuals	O
that	O
are	O
Deaf	O
or	O
Hard	O
of	O
Hearing	O
,	O
speech	B-Application
recognition	I-Application
software	I-Application
is	O
used	O
to	O
automatically	O
generate	O
a	O
closed-captioning	O
of	O
conversations	O
such	O
as	O
discussions	O
in	O
conference	O
rooms	O
,	O
classroom	O
lectures	O
,	O
and/or	O
religious	O
services	O
.	O
</s>
<s>
Speech	B-Application
recognition	I-Application
is	O
also	O
very	O
useful	O
for	O
people	O
who	O
have	O
difficulty	O
using	O
their	O
hands	O
,	O
ranging	O
from	O
mild	O
repetitive	O
stress	O
injuries	O
to	O
involve	O
disabilities	O
that	O
preclude	O
using	O
conventional	O
computer	O
input	O
devices	O
.	O
</s>
<s>
In	O
fact	O
,	O
people	O
who	O
used	O
the	O
keyboard	O
a	O
lot	O
and	O
developed	O
RSI	O
became	O
an	O
urgent	O
early	O
market	O
for	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
Speech	B-Application
recognition	I-Application
is	O
used	O
in	O
deaf	O
telephony	O
,	O
such	O
as	O
voicemail	O
to	O
text	O
,	O
relay	O
services	O
,	O
and	O
captioned	O
telephone	O
.	O
</s>
<s>
Virtual	B-Protocol
assistant	I-Protocol
(	O
e.g.	O
</s>
<s>
The	O
performance	O
of	O
speech	B-Application
recognition	I-Application
systems	O
is	O
usually	O
evaluated	O
in	O
terms	O
of	O
accuracy	O
and	O
speed	O
.	O
</s>
<s>
Accuracy	O
is	O
usually	O
rated	O
with	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
(	O
WER	O
)	O
,	O
whereas	O
speed	O
is	O
measured	O
with	O
the	O
real	O
time	O
factor	O
.	O
</s>
<s>
Other	O
measures	O
of	O
accuracy	O
include	O
Single	O
Word	B-General_Concept
Error	I-General_Concept
Rate	I-General_Concept
(	O
SWER	O
)	O
and	O
Command	O
Success	O
Rate	O
(	O
CSR	O
)	O
.	O
</s>
<s>
Speech	B-Application
recognition	I-Application
by	O
machine	O
is	O
a	O
very	O
complex	O
problem	O
,	O
however	O
.	O
</s>
<s>
Accuracy	O
of	O
speech	B-Application
recognition	I-Application
may	O
vary	O
with	O
the	O
following	O
:	O
</s>
<s>
As	O
mentioned	O
earlier	O
in	O
this	O
article	O
,	O
the	O
accuracy	O
of	O
speech	B-Application
recognition	I-Application
may	O
vary	O
depending	O
on	O
the	O
following	O
factors	O
:	O
</s>
<s>
Speech	B-Application
recognition	I-Application
is	O
a	O
multi-leveled	O
pattern	O
recognition	O
task	O
.	O
</s>
<s>
Phonemes	B-Language
,	O
Words	O
,	O
Phrases	O
,	O
and	O
Sentences	O
;	O
</s>
<s>
By	O
combining	O
decisions	O
probabilistically	O
at	O
all	O
lower	O
levels	O
,	O
and	O
making	O
more	O
deterministic	O
decisions	O
only	O
at	O
the	O
highest	O
level	O
,	O
speech	B-Application
recognition	I-Application
by	O
a	O
machine	O
is	O
a	O
process	O
broken	O
into	O
several	O
phases	O
.	O
</s>
<s>
In	O
order	O
to	O
expand	O
our	O
knowledge	O
about	O
speech	B-Application
recognition	I-Application
,	O
we	O
need	O
to	O
take	O
into	O
consideration	O
neural	B-Architecture
networks	I-Architecture
.	O
</s>
<s>
There	O
are	O
four	O
steps	O
of	O
neural	B-Architecture
network	I-Architecture
approaches	O
:	O
</s>
<s>
Compute	O
features	O
of	O
spectral-domain	O
of	O
the	O
speech	O
(	O
with	O
Fourier	B-Algorithm
transform	I-Algorithm
)	O
;	O
</s>
<s>
Analysis	O
of	O
four-step	O
neural	B-Architecture
network	I-Architecture
approaches	O
can	O
be	O
explained	O
by	O
further	O
information	O
.	O
</s>
<s>
Basic	O
sound	O
creates	O
a	O
wave	O
which	O
has	O
two	O
descriptions	O
:	O
amplitude	B-Application
(	O
how	O
strong	O
is	O
it	O
)	O
,	O
and	O
frequency	O
(	O
how	O
often	O
it	O
vibrates	O
per	O
second	O
)	O
.	O
</s>
<s>
Accuracy	O
can	O
be	O
computed	O
with	O
the	O
help	O
of	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
(	O
WER	O
)	O
.	O
</s>
<s>
Word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
can	O
be	O
calculated	O
by	O
aligning	O
the	O
recognized	O
word	O
and	O
referenced	O
word	O
using	O
dynamic	O
string	O
alignment	O
.	O
</s>
<s>
The	O
problem	O
may	O
occur	O
while	O
computing	O
the	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
due	O
to	O
the	O
difference	O
between	O
the	O
sequence	O
lengths	O
of	O
the	O
recognized	O
word	O
and	O
referenced	O
word	O
.	O
</s>
<s>
The	O
formula	O
to	O
compute	O
the	O
word	B-General_Concept
error	I-General_Concept
rate	I-General_Concept
(	O
WER	O
)	O
is	O
:	O
</s>
<s>
While	O
computing	O
,	O
the	O
word	B-General_Concept
recognition	I-General_Concept
rate	I-General_Concept
(	O
WRR	O
)	O
is	O
used	O
.	O
</s>
<s>
Speech	B-Application
recognition	I-Application
can	O
become	O
a	O
means	O
of	O
attack	O
,	O
theft	O
,	O
or	O
accidental	O
operation	O
.	O
</s>
<s>
The	O
other	O
adds	O
small	O
,	O
inaudible	O
distortions	O
to	O
other	O
speech	O
or	O
music	O
that	O
are	O
specially	O
crafted	O
to	O
confuse	O
the	O
specific	O
speech	B-Application
recognition	I-Application
system	O
into	O
recognizing	O
music	O
as	O
speech	O
,	O
or	O
to	O
make	O
what	O
sounds	O
like	O
one	O
command	O
to	O
a	O
human	O
sound	O
like	O
a	O
different	O
command	O
to	O
the	O
system	O
.	O
</s>
<s>
Popular	O
speech	B-Application
recognition	I-Application
conferences	O
held	O
each	O
year	O
or	O
two	O
include	O
SpeechTEK	O
and	O
SpeechTEK	O
Europe	O
,	O
ICASSP	O
,	O
Interspeech/Eurospeech	O
,	O
and	O
the	O
IEEE	O
ASRU	O
.	O
</s>
<s>
Conferences	O
in	O
the	O
field	O
of	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
such	O
as	O
ACL	O
,	O
NAACL	O
,	O
EMNLP	O
,	O
and	O
HLT	O
,	O
are	O
beginning	O
to	O
include	O
papers	O
on	O
speech	B-Algorithm
processing	I-Algorithm
.	O
</s>
<s>
Books	O
like	O
"	O
Fundamentals	O
of	O
Speech	B-Application
Recognition	I-Application
"	O
by	O
Lawrence	O
Rabiner	O
can	O
be	O
useful	O
to	O
acquire	O
basic	O
knowledge	O
but	O
may	O
not	O
be	O
fully	O
up	O
to	O
date	O
(	O
1993	O
)	O
.	O
</s>
<s>
Another	O
good	O
source	O
can	O
be	O
"	O
Statistical	O
Methods	O
for	O
Speech	B-Application
Recognition	I-Application
"	O
by	O
Frederick	O
Jelinek	O
and	O
"	O
Spoken	O
Language	O
Processing	O
(	O
2001	O
)	O
"	O
by	O
Xuedong	O
Huang	O
etc.	O
,	O
"	O
Computer	O
Speech	O
"	O
,	O
by	O
Manfred	O
R	O
.	O
Schroeder	O
,	O
second	O
edition	O
published	O
in	O
2004	O
,	O
and	O
"	O
Speech	B-Algorithm
Processing	I-Algorithm
:	O
A	O
Dynamic	O
and	O
Optimization-Oriented	O
Approach	O
"	O
published	O
in	O
2003	O
by	O
Li	O
Deng	O
and	O
Doug	O
O'Shaughnessey	O
.	O
</s>
<s>
Speaker	B-Application
recognition	I-Application
also	O
uses	O
the	O
same	O
features	O
,	O
most	O
of	O
the	O
same	O
front-end	O
processing	O
,	O
and	O
classification	O
techniques	O
as	O
is	O
done	O
in	O
speech	B-Application
recognition	I-Application
.	O
</s>
<s>
A	O
comprehensive	O
textbook	O
,	O
"	O
Fundamentals	O
of	O
Speaker	B-Application
Recognition	I-Application
"	O
is	O
an	O
in	O
depth	O
source	O
for	O
up	O
to	O
date	O
details	O
on	O
the	O
theory	O
and	O
practice	O
.	O
</s>
<s>
A	O
good	O
insight	O
into	O
the	O
techniques	O
used	O
in	O
the	O
best	O
modern	O
systems	O
can	O
be	O
gained	O
by	O
paying	O
attention	O
to	O
government	O
sponsored	O
evaluations	O
such	O
as	O
those	O
organised	O
by	O
DARPA	O
(	O
the	O
largest	O
speech	O
recognition-related	O
project	O
ongoing	O
as	O
of	O
2007	O
is	O
the	O
GALE	O
project	O
,	O
which	O
involves	O
both	O
speech	B-Application
recognition	I-Application
and	O
translation	O
components	O
)	O
.	O
</s>
<s>
A	O
good	O
and	O
accessible	O
introduction	O
to	O
speech	B-Application
recognition	I-Application
technology	I-Application
and	O
its	O
history	O
is	O
provided	O
by	O
the	O
general	O
audience	O
book	O
"	O
The	O
Voice	O
in	O
the	O
Machine	O
.	O
</s>
<s>
The	O
most	O
recent	O
book	O
on	O
speech	B-Application
recognition	I-Application
is	O
Automatic	B-Application
Speech	I-Application
Recognition	I-Application
:	O
A	O
Deep	B-Algorithm
Learning	I-Algorithm
Approach	O
(	O
Publisher	O
:	O
Springer	O
)	O
written	O
by	O
Microsoft	O
researchers	O
D	O
.	O
Yu	O
and	O
L	O
.	O
Deng	O
and	O
published	O
near	O
the	O
end	O
of	O
2014	O
,	O
with	O
highly	O
mathematically	O
oriented	O
technical	O
detail	O
on	O
how	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
are	O
derived	O
and	O
implemented	O
in	O
modern	O
speech	B-Application
recognition	I-Application
systems	O
based	O
on	O
DNNs	O
and	O
related	O
deep	B-Algorithm
learning	I-Algorithm
methods	O
.	O
</s>
<s>
A	O
related	O
book	O
,	O
published	O
earlier	O
in	O
2014	O
,	O
"	O
Deep	B-Algorithm
Learning	I-Algorithm
:	O
Methods	O
and	O
Applications	O
"	O
by	O
L	O
.	O
Deng	O
and	O
D	O
.	O
Yu	O
provides	O
a	O
less	O
technical	O
but	O
more	O
methodology-focused	O
overview	O
of	O
DNN-based	O
speech	B-Application
recognition	I-Application
during	O
2009	O
–	O
2014	O
,	O
placed	O
within	O
the	O
more	O
general	O
context	O
of	O
deep	B-Algorithm
learning	I-Algorithm
applications	O
including	O
not	O
only	O
speech	B-Application
recognition	I-Application
but	O
also	O
image	O
recognition	O
,	O
natural	B-Language
language	I-Language
processing	I-Language
,	O
information	O
retrieval	O
,	O
multimodal	O
processing	O
,	O
and	O
multitask	O
learning	O
.	O
</s>
<s>
In	O
terms	O
of	O
freely	O
available	O
resources	O
,	O
Carnegie	O
Mellon	O
University	O
's	O
Sphinx	B-General_Concept
toolkit	O
is	O
one	O
place	O
to	O
start	O
to	O
both	O
learn	O
about	O
speech	B-Application
recognition	I-Application
and	O
to	O
start	O
experimenting	O
.	O
</s>
<s>
Another	O
resource	O
(	O
free	O
but	O
copyrighted	O
)	O
is	O
the	O
HTK	B-General_Concept
book	O
(	O
and	O
the	O
accompanying	O
HTK	B-General_Concept
toolkit	O
)	O
.	O
</s>
<s>
For	O
more	O
recent	O
and	O
state-of-the-art	O
techniques	O
,	O
Kaldi	B-General_Concept
toolkit	O
can	O
be	O
used	O
.	O
</s>
<s>
In	O
2017	O
Mozilla	B-Operating_System
launched	O
the	O
open	O
source	O
project	O
called	O
Common	B-General_Concept
Voice	I-General_Concept
to	O
gather	O
big	O
database	O
of	O
voices	O
that	O
would	O
help	O
build	O
free	O
speech	B-Application
recognition	I-Application
project	O
DeepSpeech	O
(	O
available	O
free	O
at	O
GitHub	B-Application
)	O
,	O
using	O
Google	B-Application
's	I-Application
open	O
source	O
platform	O
TensorFlow	B-Language
.	O
</s>
<s>
When	O
Mozilla	B-Operating_System
redirected	O
funding	O
away	O
from	O
the	O
project	O
in	O
2020	O
,	O
it	O
was	O
forked	O
by	O
its	O
original	O
developers	O
as	O
Coqui	O
STT	O
using	O
the	O
same	O
open-source	O
license	O
.	O
</s>
<s>
Google	B-Application
Gboard	B-Device
supports	O
speech	B-Application
recognition	I-Application
on	O
all	O
Android	B-Application
applications	I-Application
.	O
</s>
<s>
It	O
can	O
be	O
activated	O
through	O
the	O
microphone	B-Application
icon	O
.	O
</s>
<s>
The	O
commercial	O
cloud	O
based	O
speech	B-Application
recognition	I-Application
APIs	O
are	O
broadly	O
available	O
.	O
</s>
<s>
For	O
more	O
software	O
resources	O
,	O
see	O
List	B-General_Concept
of	I-General_Concept
speech	I-General_Concept
recognition	I-General_Concept
software	I-General_Concept
.	O
</s>
