Included models¶
This distribution includes sample models in ~/Sensory/TrulyNaturalSDK/7.9.0/model/
You can also download models for additional languages in a range of sizes.
Console examples in this section assume $HOME/Sensory/TrulyNaturalSDK/7.9.0 as the SDK install directory; replace that prefix if you installed elsewhere.
Wake words¶
See the wake word model type for description of model behavior and settings.
spot-voicegenie-enUS-6.6.0-m.snsr¶
Fixed-phrase "Voice Genie" wake word for US English. Reports a confidence-score with each detection.
spot-hbg-enUS-1.4.0-m.snsr¶
Fixed-phrase "Hello Blue Genie" wake word for US English.
spot-music-enUS-1.2.0-m.snsr¶
Music command set for US English. Commands include "play music", "pause music", "stop music", "previous song", and "next song".
Adapting wake word¶
See the adapting wake word model type for a description of model behavior and settings.
ca-voicegenie-enUS-1.1.0.snsr¶
This is a fixed-phrase spotter for "Voice Genie" in US English that adapts to users' speech to improve false-accept rates.
Model adaptation and enrollment happens automatically and without any additional code requirements — you can use this model as a drop-in replacement for the fixed-phrase Voice Genie spotter.
Configuration settings of particular interest include cache-file and max-users.
Use user-iterator, delete-user, and rename-user to manage user enrollments.
Note
This model requires support for multi-threading.
Example
The reported text value changes once enrollment has identified and enrolled a new speaker.
% cd $HOME/Sensory/TrulyNaturalSDK/7.9.0
# two different speakers saying "voice genie"
% bin/snsr-eval -t model/ca-voicegenie-enUS-1.1.0.snsr\
-s cache-file=ca-vg-cache.snsr\
-s max-users=3
2235 3045 voice_genie
6810 7545 voice_genie
11745 12525 user1/voice_genie
16845 17595 user1/voice_genie
29355 30180 voice_genie
34845 35820 user2/voice_genie
37815 38520 user1/voice_genie
40080 40905 user2/voice_genie
^C
# restart, loading enrollments from the cache file
% bin/snsr-eval -t model/ca-voicegenie-enUS-1.1.0.snsr\
-s cache-file=ca-vg-cache.snsr\
-s max-users=3
12045 13035 user2/voice_genie
15180 15840 user1/voice_genie
17745 18465 user1/voice_genie
20175 20820 user2/voice_genie
^C
Wake word enrollment¶
See the wake word enrollment model type for a description of model behavior and settings.
eft-hbg-enUS-23.0.0.9.snsr¶
EFT spotter for "Hello Blue Genie", US English. This model produces wake words with a low imposter accept rate.
udt-universal-3.67.1.0.snsr¶
UDT enrollment. This model creates spotters with nine different operating points and supports multiple languages.
Optimized for German, English (Australian, British, Indian, United States), Spanish (European Union, North American), French (European Union), Italian, Korean, Brazilian Portuguese, and Mandarin Chinese.
udt-enUS-5.1.1.9.snsr¶
UDT enrollment with backwards compatibility.
Note
This older model produces enrolled wake words with reduced accuracy.
Use this model only when targeting a THF Micro 3.x DSP port, or when the wake word is followed by additional validation.
VAD¶
See the VAD model type for a description of model behavior and settings.
vad-ml-3.0.0.snsr¶
Deep-learned stand-alone Voice Activity Detector.
LVCSR tnl¶
See the LVCSR model type for a description of model behavior and settings.
lvcsr-build-enUS-14.2.0-5MB.snsr tnl¶
US English recognizer with 4.9 MiB acoustic model and support for grammar-based recognition and in-grammar acoustic confidence scores. Use search.frame-nota to adjust out-of-grammar rejection.
Example
% cd $HOME/Sensory/TrulyNaturalSDK/7.9.0
% snsr-eval -t model/lvcsr-build-enUS-14.2.0-5MB.snsr\
-s partial-result-interval=0 \
-f grammar-stream data/grammars/en-US/enrollments-nlu-slot.grm \
data/enrollments/armadillo-1-2-c.wav
NLU intent: navigate (0.0000) = how far away is winco
NLU entity: place (0.0000) = winco
300 1920 armadillo how far away is winco
STT stt¶
See the STT model type for a description of model behavior and settings.
You can download additional models for other languages in a range of different sizes.
stt-enUS-general-medium-2.4.5-pnc.snsr stt¶
STT recognizer with broad-domain and grammar-based recognition support. Results include capitalization and punctuation.
This model requires STT support, which currently depends on third-party Open Source modules that are optionally included in the TrulyNatural SDK See Open Source Licenses for details.
Templates¶
Templates add functionality to recognizer models. This includes running models simultaneously or sequentially, and adding VAD audio gating.
See template types for an overview.
tpl-spot-concurrent-1.6.0.snsr¶
Runs two wake word models at the same time.
tpl-spot-debug-1.6.0.snsr¶
Adds runtime data collection to a wake word model.
tpl-spot-select-1.6.0.snsr¶
Dynamically selects which of the two embedded wake word models to run.
tpl-spot-sequential-1.6.0.snsr¶
Runs two wake word models one after the other, with optional looping on the second. Includes push-to-talk as an alternative to the wake word.
tpl-spot-vad-3.14.0.snsr¶
Runs a wake word until it spots, then does start- and endpoint detection on the subsequent audio stream using a VAD.
tpl-opt-spot-vad-lvcsr-1.29.0.snsr tnl¶
Optionally runs a wake word until it spots, segments the subsequent audio stream with a VAD, then sends the segmented audio to an LVCSR or STT recognizer. You can select at runtime whether recognition waits on the wake word or starts immediately.
tpl-spot-vad-lvcsr-3.24.0.snsr tnl¶
Runs a wake word until it spots, segments the subsequent audio stream with a VAD, then sends the segmented audio to an LVCSR or STT recognizer.
tpl-vad-lvcsr-3.18.0.snsr¶
Detects speech with a VAD and sends the segmented audio to an LVCSR or STT recognizer.