Grammar structures for smart homes dialogues

Download 0,56 Mb. Pdf ko'rish
bet	9/14
Sana	15.05.2024
Hajmi	0,56 Mb.
	#235098

1 ... 6 7 8 9 10 11 12 13 14

6. Grammar structures for smart homes dialogues
As seen in section 4, VI for smart homes involves different levels of complexity ranging from
an isolated word recognizer to an unconstrained ASR dictation system. For isolated word
recognition, a speech recognizer attempts to detect commands by focusing on keywords in
segmented natural sentences from the talker. A more complicated system, often referred to as
a system driven dialogue (Nakano et al., 2006), can allow a user to complete information fields.
Large vocabulary continuous speech recognition (LVCSR), by contrast, requires language
models or grammars to select the most likely word sequence from the relatively large
number of alternative word hypotheses produced during the search process at each stage in
a conversation. Simple recognition tasks can use rule-based regular or context free
grammars, where the system only recognizes a limited vocabulary.
In a speech-based smart home system, a VI session begins with a configurable spoken
identifier phrase to 'attract the attention' of the ASR computer. From this point on, the
adoption of grammar and syntactic structures, although these must be learnt by the user
and thus reduce user friendliness, are crucial in maintaining the recognition accuracy of
such systems.
Following attention, a tree of vocabulary options is possible. The aim at each stage of is to
maximise recognition accuracy for given SNR, by reducing the vocabulary search space. An
example sentence for one speaker might be:
ACTION – argument – (optional MODIFIER- (optional argument)) – PAUSE/REPEAT
In general, each of the action, argument and modifier sub phrases will have differing
vocabulary characteristics, but it is a common feature of such systems that vocabulary size
be minimised through syntactic structure. Let us proposed that each sub-phrase is classified
by the following arguments:
•
Vocabulary consisting of V elements
•
Accuracy requirement, R
•
Length constraint, L
•
Interruptibility type, T
The characteristics of each sub-phrase are known in advance, having been defined by the
grammar syntax, with recall prompted by traversal to the current branch. The length
constraint is used to help detect end-of-phrase overrun conditions, and perhaps a missed
MODIFIER phrase.
One or more phrases by each party in the communication comprise a conversation (or
dialogue). An overall conversation, and indeed a phrase made of sub phrases has its own
accuracy requirement. However this level of end-to-end link assurance is unwieldy for
anything but confirming that an important action should or should not occur as the result of
a conversation.
Sub-phrase vocabulary may occasionally need to be unconstrained (such as when
performing a web search or dictating an email), but at other times could be limited (in this
mode, we can use keywords or embedded phrases in natural sentences to command the

Speech Recognition for Smart Homes
487
system). Potentially these two classes of recognition task could be performed with two
different types of ASR software employing isolated word and continuous speech
recognition respectively (Nakano et al., 2006). However both are catered for in Sphinx2
using two density acoustic models, namely semi-continuous and continuous. Some other
flexible speech recognition systems have been introduced.
One system by Furui (Furui, 2001) uses a method of automatically summarizing speech,
sentence by sentence. This is quite domain-specific (closed-domain) with limited vocabulary
size (designed for news transcription), and may not be directly applicable to a continuously
variable vocabulary size smart home system. However it does perform very well, and
provides an indication of the customisations available in such systems.
Vocabulary size,
V
, impacts recognition accuracy, and needs to be related to accuracy
requirement,
R
. Since 100% accuracy is unlikely, smart home systems need to be able to cope
with inaccuracies through careful design of the interaction and confirmation processes. In

Download 0,56 Mb.

1 ... 6 7 8 9 10 11 12 13 14

Download 0,56 Mb.

Pdf ko'rish