pylon / spokestack-android

91%

master: 91%

LAST BUILD ON BRANCH bs-wakeword-2

branch: bs-wakeword-2

CHANGE BRANCH

x

Reset

• bs-wakeword-2
• bs-2.0
• bs-2.0.1
• bs-bing-asr
• bs-circle-badge
• bs-circleci
• bs-google-asr
• bs-manual-deactivate
• bs-noise-suppressor
• bs-pipeline
• bs-release-0.2.2
• bs-timeout
• bs-trace
• bs-wakeword
• jz-migration
• master
• ndw-0.1.8
• ndw-RejectedExecutionException
• ndw-deps
• ndw-grpc
• ndw-release
• ndw-time-out

add support for autoregressive detectors

These changes introduce a new tensorflow "encode" model in the
wakeword trigger, which allows us to implement autoregressive models
for wakeword detection (like rnns and attention). Doing this
efficiently requires running each frame through the encoder and then
aggregating these frames for attention/classification, in order to
avoid redundant calls to the encoder. These models also provide the
ability to detect whole phrases with long-range frame dependencies,
so we no longer need manual phrasing or smoothing logic and buffers.
For an example model, see https://arxiv.org/abs/1803.10916