BelaSigna R281
www.onsemi.com
6
Description of Operation
Efficient implementation of the trigger phrase recognition
algorithm is accomplished through the use of three
processing units running concurrently. In addition to the
main DSP core performing system configuration and signal
processing, an input/output processor continuously collects
frames of input speech signals which are analyzed by a
highly optimized frequency domain co−processor.
There are two main modes of operation: Recognition
Mode and Training Mode.
Recognition Mode
When in Recognition Mode, the entire system remains in
an extremely low−power, “always−on” state continuously
listening for speech. When speech is detected, the algorithm
proceeds to extract features from the collected audio data,
comparing these features to a known set of data computed
during the training process. If the features appear to be
similar enough to the trained feature data, a match is
indicated on the WAKE_UP pin.
Once the feature extraction and comparison is complete,
the system returns to its original low−power state, once
again waiting until speech is detected.
Training Mode
Before BelaSigna R281 can be placed into Recognition
Mode, it must be trained. Training involves recording and
analyzing three utterances of a trigger phrase. A trigger
phrase can be any collection of words or sounds, in any
language, up to a maximum length of approximately 1.5
seconds.
When speaking a given phrase, a human will naturally say
the same phrase subtly different each time. Having multiple
instances of the same phrase, each slightly different, helps
to make the matching algorithm more robust.
Training must be performed in a quiet environment, or the
trigger phrase match results in Recognition Mode will be
unpredictable. BelaSigna R281 is placed into Training
Mode by issuing an appropriate SetMode command via I
2
C.
The algorithm stores relevant feature data from the three
separate utterances of the trigger phrase, also known as
training templates.
Once training is complete, the training templates can be
read from memory and stored offline. This template data can
then be loaded into memory in the event of a power cycle and
the device can be placed directly into Recognition Mode,
thus avoiding the need to re−train the device. This same
procedure can be used to recognize multiple trigger phrases,
provided they have each been individually trained and
stored offline. Only one trigger phrase can be active at any
one time.
Refer to AND9267/D for more information on the
different modes, as well as a description of the I
2
C host
control protocol.
Initial Power−On State
When BelaSigna R281 is powered on, the device will
perform a brief initialization procedure and then wait for a
connection to be made from an external host via I
2
C. At this
point, the host controller must connect to BelaSigna R281
and load its memory with the algorithm binary image, as
well as the training template data. Once this has been
completed, the device can be put into Recognition Mode. If
no training template data is available (e.g. the training
procedure has never been performed), then BelaSigna R281
must be placed into Training Mode and the training
procedure performed before entering Recognition Mode.
Whenever power is removed from the device, the contents
of memory are lost and must be re−loaded.
For more information refer to AND9267/D.
