--- source_path: "getting-started/index.md" canonical_url: "https://doc.sensory.com/tnl/7.8/getting-started/" --- # Quick start This walk-through helps you get started with TrulyNatural concepts and the command-line tools (`snsr-eval`, `snsr-edit`, and related utilities). **Ready to embed in your own app?:** See [Your first program](https://doc.sensory.com/tnl/7.8/getting-started/your-first-program.md#your-first-program) for a hands-on Session API walk-through with platform tabs for **C** (CMake), **Java** (Gradle), **Python** ([uv][] and the SDK wheel — [Integrate with your build § Python](https://doc.sensory.com/tnl/7.8/api/build-system.md#build-python)), **Android**, and **iOS** (C API from Swift via a bridging header). **Looking for sample code?:** See [C](https://doc.sensory.com/tnl/7.8/api/sample/c/index.md#c-examples), [Java](https://doc.sensory.com/tnl/7.8/api/sample/java/index.md#java-examples), [Python](https://doc.sensory.com/tnl/7.8/api/sample/python/index.md#python-examples), [Android](https://doc.sensory.com/tnl/7.8/api/sample/android/index.md#android-examples), and [iOS](https://doc.sensory.com/tnl/7.8/api/sample/ios/index.md#ios-examples) in the [API reference](https://doc.sensory.com/tnl/7.8/api/index.md#api-reference) section. ## Prerequisites The paths below assume the default online documentation install location, _$HOME/Sensory/TrulyNaturalSDK/7.9.0-pre.0_. If you installed the SDK elsewhere, replace that prefix with your SDK installation directory. Offline SDK documentation uses the actual installed path. 1. Install the TrulyNatural SDK using the provided installer executable. 2. Open a terminal with a command-line prompt. 3. Add _$HOME/Sensory/TrulyNaturalSDK/7.9.0-pre.0/bin_ to your shell `PATH` variable, and change your working directory to _$HOME/Sensory/TrulyNaturalSDK/7.9.0-pre.0_ **Configure your shell environment** ```shell PATH=${PATH}:$HOME/Sensory/TrulyNaturalSDK/7.9.0-pre.0/bin cd $HOME/Sensory/TrulyNaturalSDK/7.9.0-pre.0 ``` ## Wake words 1. Start with running a simple wake word on live audio input. The [snsr-eval](https://doc.sensory.com/tnl/7.8/tools/snsr-eval.md#snsr-eval) utility can run all recognition and VAD model types. Use the `-t` (_task_) flag to specify the path to the wake word model. Start `snsr-eval` and say "voice genie" a number of times: ```console % snsr-eval -t model/spot-voicegenie-enUS-6.5.1-m.snsr 1275 1920 voicegenie 5070 5730 voicegenie 10395 10980 voicegenie ^C ``` The output shows the start and end times in ms since the start of the recording, and the phrase the key word spotter detected. `snsr-eval` runs until you interrupt it with `^C`. The [spot-voicegenie-enUS-6.5.1-m.snsr](https://doc.sensory.com/tnl/7.8/models/index.md#spot-voicegenie-enUS) model file includes everything needed to run this wake word, including reasonable default settings. For wake words, there is one configuration option that you might want to adjust: [operating-point](https://doc.sensory.com/tnl/7.8/api/setting-keys/configuration.md#operating-point), which controls the recognition sensitivity. You can do this on the command-line with the `-s` option: ```console % snsr-eval -t model/spot-voicegenie-enUS-6.5.1-m.snsr -s operating-point=21 1080 1605 voicegenie 2610 3180 voicegenie 5460 5970 voicegenie ^C ``` Increase [snsr-eval](https://doc.sensory.com/tnl/7.8/tools/snsr-eval.md#snsr-eval) output verbosity with one or more `-v` flags: ```console % snsr-eval -vt model/spot-voicegenie-enUS-6.5.1-m.snsr Using live audio from default capture device. ^C to stop. 1410 1995 (0.999) voicegenie ^C % snsr-eval -vvt model/spot-voicegenie-enUS-6.5.1-m.snsr Using live audio from default capture device. ^C to stop. Using operating point 8. Available operating points: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21. Available vocabulary: 1: "voicegenie" phrase: 495 1020 (0.9951) voicegenie words: 495 1020 (0.9951) voicegenie ^C ``` The value between the parentheses is the recognition [score](https://doc.sensory.com/tnl/7.8/api/setting-keys/results.md#score). You can also reduce the output verbosity with one or more `-l` flags. `snsr-eval -ll` reports only the recognition result, which makes it suitable for scripted batch testing. **Note:** type: tip Run [snsr-eval](https://doc.sensory.com/tnl/7.8/tools/snsr-eval.md#snsr-eval), or any of the [command-line tools](https://doc.sensory.com/tnl/7.8/tools/index.md#command-line-tools) without arguments to see a brief usage summary and a list of available options. 2. To recognize pre-recorded audio, specify an audio file in either [RIFF WAV][] format or as a binary file containing audio samples and no header. Most models require 16-bit LPCM encoding sampled at 16 kHz. ```console % snsr-eval -t model/spot-voicegenie-enUS-6.5.1-m.snsr \ data/audio/voice-genie-set-cruise-control.wav 2310 2910 voicegenie ``` `snsr-eval` ends once it has processed the entire file. If you specify multiple files `snsr-eval` concatenates the files in order and evaluates the model on the result. Recognition timestamps reflect this concatenation. ```console % snsr-eval -t model/spot-voicegenie-enUS-6.5.1-m.snsr \ data/audio/voice-genie-set-cruise-control.wav \ data/audio/voice-genie-set-cruise-control.wav 2310 2910 voicegenie 9075 9645 voicegenie ``` 3. [snsr-eval](https://doc.sensory.com/tnl/7.8/tools/snsr-eval.md#snsr-eval) also runs [command sets](https://doc.sensory.com/tnl/7.8/faq.md#use-command-set). These are keyword spotters with more than one active word or phrase, optimized for a low false reject rate. Inspect the [sample music control](https://doc.sensory.com/tnl/7.8/models/index.md#spot-music-enUS) model vocabulary: ```console % snsr-eval -vvt model/spot-music-enUS-1.2.0-m.snsr /dev/null Using operating point 17. Available operating points: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20. Available vocabulary: 1: "play_music" 2: "previous_song" 3: "stop_music" 4: "next_song" 5: "pause_music" ``` Command sets work just like wake words: ```console % snsr-eval -t model/spot-music-enUS-1.2.0-m.snsr \ data/audio/voice-genie-music.wav 5160 5865 play_music 8055 8790 next_song 14820 15705 stop_music ``` 4. Running an [adapting wake word](https://doc.sensory.com/tnl/7.8/models/types/ca.md#ca-type) uses the same recipe as regular wake words: ```console % snsr-eval -vt model/ca-voicegenie-enUS-1.1.0.snsr Using live audio from default capture device. ^C to stop. [^adapt-started] on worker thread 8295 8790 (1 sv) voice_genie [^adapt-started] on worker thread 15645 16245 (1 sv) voice_genie 16515 [^adapted] 16515 [^new-user] user1/voice_genie [^adapt-started] on worker thread 20445 21060 (0.9678 sv) user1/voice_genie 21225 [^adapted] ^C ``` Note how the result changes once the model has adapted to your speech. If you have a second speaker say "voice genie" a couple of times you should see `user2/voice_genie` for their utterances. The model adaptations do not persist and will be lost when the model is reloaded. To address that, specify a [cache-file](https://doc.sensory.com/tnl/7.8/api/setting-keys/configuration.md#cache-file) to hold these. Say "voice genie" a number of times, then stop `snsr-eval` with `^C`: ```console % snsr-eval -vt model/ca-voicegenie-enUS-1.1.0.snsr \ -s cache-file=voice-genie.cache Using live audio from default capture device. ^C to stop. [^adapt-started] on worker thread 1485 2070 (1 sv) voice_genie [^adapt-started] on worker thread 4740 5355 (1 sv) voice_genie 5460 [^adapted] 5460 [^new-user] user1/voice_genie [^adapt-started] on worker thread 9285 9855 (0.9951 sv) user1/voice_genie 10140 [^adapted] ^C ``` `snsr-eval` loads _voice-genie.cache_ when restarting. The very first result already includes the `user1/voice_genie` result: ```console % snsr-eval -vt model/ca-voicegenie-enUS-1.1.0.snsr \ -s cache-file=voice-genie.cache Using live audio from default capture device. ^C to stop. [^adapt-started] on worker thread 1185 1815 (0.9971 sv) user1/voice_genie 2085 [^adapted] ^C ``` ## Wake word enrollment EFT models adapt a specific wake word phrase to a speaker's voice. UDT models create new wake words for arbitrary phrases from a handful of speaker-specific recordings. 1. Start with UDT. The [spot-enroll](https://doc.sensory.com/tnl/7.8/tools/spot-enroll.md#spot-enroll) utility does EFT and UDT enrollment from recordings. Most models require four recordings of the same phrase for optimal performance. UDT models are less likely to trigger on the same phrase said by another speaker. **Enroll three phrases into a single model** ```console % spot-enroll -vt model/udt-universal-3.67.1.0.snsr \ -o udt-kws.snsr \ +armadillo-1 data/enrollments/armadillo-1-{0,1,2,3}.wav \ +armadillo-6 data/enrollments/armadillo-6-{0,1,2,3}.wav \ +jackalope-1 data/enrollments/jackalope-1-{0,1,2,3}.wav \ +jackalope-4 data/enrollments/jackalope-4-{0,1,2,3}.wav \ +terminator-2 data/enrollments/terminator-2-{0,1,2,3}.wav \ +terminator-6 data/enrollments/terminator-6-{0,1,2,3}.wav Adapting: 100% complete. Enrolled model saved to "udt-kws.snsr" ``` The new `udt-kws.snsr` wake word model spots three different phrases with two unique speakers per phrase. Run the model with [snsr-eval](https://doc.sensory.com/tnl/7.8/tools/snsr-eval.md#snsr-eval), using different recordings than the ones used for enrollment: ```console % snsr-eval -t udt-kws.snsr \ data/enrollments/armadillo-1-0-c.wav \ data/enrollments/armadillo-6-0.wav \ data/enrollments/jackalope-1-0-c.wav \ data/enrollments/jackalope-4-0.wav \ data/enrollments/terminator-2-0.wav \ data/enrollments/terminator-6-0.wav 330 945 armadillo-1 4485 5265 armadillo-6 6795 7320 jackalope-1 10365 10890 jackalope-4 13245 13950 terminator-2 15510 16215 terminator-6 ``` The model identifies the phrases and the speakers correctly. 2. EFT models use the same tools and API for enrollment as UDT. This example uses [live-enroll](https://doc.sensory.com/tnl/7.8/tools/live-enroll.md#live-enroll), the interactive version of [spot-enroll](https://doc.sensory.com/tnl/7.8/tools/spot-enroll.md#spot-enroll), the main difference being that it prompts to repeat recordings that aren't usable instead of reporting an enrollment error. Run the example below, then say "hello blue genie" when prompted. ```console % live-enroll -vt model/eft-hbg-enUS-23.0.0.9.snsr \ -o eft-hbg.snsr +user-1 Say the enrollment phrase (1/4) for "user-1" Recording: 4.29 s Preliminary enrollment checks passed. Say the enrollment phrase (2/4) for "user-1" Recording: 3.23 s Preliminary enrollment checks passed. Say the enrollment phrase (3/4) for "user-1" with context, for example: " will it rain tomorrow?" Recording: 3.15 s Preliminary enrollment checks passed. Say the enrollment phrase (4/4) for "user-1" with context, for example: " will it rain tomorrow?" Recording: 3.90 s Preliminary enrollment checks passed. Adapting: 100% complete. Enrolled model saved to "eft-hbg.snsr" ``` As before, test with [snsr-eval](https://doc.sensory.com/tnl/7.8/tools/snsr-eval.md#snsr-eval): ```console % snsr-eval -t eft-hbg.snsr Using live audio from default capture device. ^C to stop. 675 1500 (0.8921 sv) user-1/HBG 3480 4500 (0.8245 sv) user-1/HBG 7020 8160 (0.877 sv) user-1/HBG ^C ``` The value between the parentheses is the [sv-score](https://doc.sensory.com/tnl/7.8/api/setting-keys/results.md#sv-score). ## Templates [Templates](https://doc.sensory.com/tnl/7.8/models/tpl/index.md#template-type) are models that add new behaviors to [wake word](https://doc.sensory.com/tnl/7.8/models/types/wake-word.md#wake-word-type), [lvcsr](https://doc.sensory.com/tnl/7.8/models/types/lvcsr.md#lvcsr-type), and [stt](https://doc.sensory.com/tnl/7.8/models/types/stt.md#stt-type) models via composition. Templates have _slots_ that you fill with models of the required type. 1. Suppose you would like to reduce false accepts of the music command set used above by requiring commands in it to be prefixed with a low false-accept wake word like "voice genie". The [tpl-spot-sequential](https://doc.sensory.com/tnl/7.8/models/tpl/tpl-spot-sequential.md#tpl-spot-sequential-type) template makes this possible: ```console % snsr-eval -t model/tpl-spot-sequential-1.5.0.snsr \ -f 0 model/spot-voicegenie-enUS-6.5.1-m.snsr \ -f 1 model/spot-music-enUS-1.2.0-m.snsr \ data/audio/voice-genie-music.wav 5175 5865 play_music ``` `snsr-eval`'s `-f slot filename` option loads the named file into the specified slot. The [tpl-spot-sequential](https://doc.sensory.com/tnl/7.8/models/tpl/tpl-spot-sequential.md#tpl-spot-sequential-type) documentation lists the slots the template supports, and the types of models it expects for these. As expected, the new recognizer spots only the command directly following "voice genie". [snsr-eval](https://doc.sensory.com/tnl/7.8/tools/snsr-eval.md#snsr-eval) supports on-the-fly model composition, but what if you have code that already works with _spot-music-enUS-1.2.0-m.snsr_ that you don't want to modify? Enter [snsr-edit](https://doc.sensory.com/tnl/7.8/tools/snsr-edit.md#snsr-edit), which supports composition and [setting](https://doc.sensory.com/tnl/7.8/api/setting-keys/index.md#setting-keys) changes and can save the result as a new, self-contained model: ```console % snsr-edit -vvt model/tpl-spot-sequential-1.5.0.snsr \ -o vg-music.snsr \ -f 0 model/spot-voicegenie-enUS-6.5.1-m.snsr \ -f 1 model/spot-music-enUS-1.2.0-m.snsr Loading "model/tpl-spot-sequential-1.5.0.snsr" as the template model. Loading "model/spot-voicegenie-enUS-6.5.1-m.snsr" into setting "0". Loading "model/spot-music-enUS-1.2.0-m.snsr" into setting "1". Output written to "vg-music.snsr". % snsr-eval -t vg-music.snsr data/audio/voice-genie-music.wav 5175 5865 play_music ``` [tpl-spot-sequential](https://doc.sensory.com/tnl/7.8/models/tpl/tpl-spot-sequential.md#tpl-spot-sequential-type) has a [loop](https://doc.sensory.com/tnl/7.8/api/setting-keys/configuration.md#loop) setting that changes behavior. Try it: ```console % snsr-eval -t vg-music.snsr -s loop=1 data/audio/voice-genie-music.wav 5175 5865 play_music 8055 8790 next_song ``` This recognizes the first two music commands, but not "stop music" as the gap between "next song" and "stop music" is more than the [listen-window](https://doc.sensory.com/tnl/7.8/api/setting-keys/configuration.md#listen-window), which is five seconds: ``` % snsr-edit -t model/spot-music-enUS-1.2.0-m.snsr -q listen-window listen-window = 5 ``` 2. Run two keyword spotters simultaneously with [tpl-spot-concurrent](https://doc.sensory.com/tnl/7.8/models/tpl/tpl-spot-concurrent.md#tpl-spot-concurrent-type): ```console % snsr-eval -t model/tpl-spot-concurrent-1.5.0.snsr \ -f 0 model/spot-voicegenie-enUS-6.5.1-m.snsr \ -f 1 model/spot-music-enUS-1.2.0-m.snsr \ data/audio/voice-genie-music.wav 4485 5085 voicegenie 5160 5865 play_music 8055 8790 next_song 14820 15705 stop_music ``` ## Speech To Text _(STT only)_ TrulyNatural STT includes support for modern transformer-based end-to-end recognizers suitable for transcription tasks. 1. STT models use the same tools and API as wake words. Run a sample audio file through [stt-enUS-automotive-medium-2.3.15-pnc.snsr ](https://doc.sensory.com/tnl/7.8/models/index.md#stt-enUS-automotive-medium) with [snsr-eval](https://doc.sensory.com/tnl/7.8/tools/snsr-eval.md#snsr-eval): ```console % snsr-eval -t model/stt-enUS-automotive-medium-2.3.15-pnc.snsr \ data/audio/voice-genie-set-cruise-control.wav P 2040 2960 God Boice jeni P 2240 3560 Voice. Genie said the creek P 2240 3960 Voice. Genie set the cruise control P 2240 4200 Voice. Genie set the cruise control to P 2240 4720 Voice. Genie set the cruise control to further P 2240 5200 Voice. Genie set the cruise control to fifty five. Mrs P 2240 5600 Voice Genie set the cruise control to fifty five miles back P 2240 5800 Voice Genie set the cruise control to fifty five miles per hour P 2240 5840 Voice Genie set the cruise control to fifty five miles per hour NLU intent: set_cruise_control (0.9969) = voice genie set the cruise control to 55 miles per hour NLU entity: number (0.9931) = 55 NLU entity: speed_unit (0.9942) = miles per hour 2240 5840 Voice Genie set the cruise control to fifty five miles per hour. ``` Partial or interim hypotheses are shown prefixed with `P`. These provide useful feedback for live transcription tasks, but are less interesting when recognizing from file. You can suppress them by setting [partial-result-interval](https://doc.sensory.com/tnl/7.8/api/setting-keys/configuration.md#partial-result-interval) `= 0`: ```console % snsr-eval -t model/stt-enUS-automotive-medium-2.3.15-pnc.snsr \ -s partial-result-interval=0 \ data/audio/voice-genie-set-cruise-control.wav NLU intent: set_cruise_control (0.9969) = voice genie set the cruise control to 55 miles per hour NLU entity: number (0.9931) = 55 NLU entity: speed_unit (0.9942) = miles per hour 2240 5840 Voice Genie set the cruise control to fifty five miles per hour. ``` 2. STT and LVCSR models only produce a final recognition hypothesis at end-of-file, or when a VAD signals that speech has ended. For convenience [snsr-eval](https://doc.sensory.com/tnl/7.8/tools/snsr-eval.md#snsr-eval) has an `-a` option that adds the [tpl-vad-lvcsr](https://doc.sensory.com/tnl/7.8/models/tpl/tpl-vad-lvcsr.md#tpl-vad-lvcsr-type) template if you're using live audio and specify an STT model that does not include a VAD. ```console % snsr-eval -lt model/stt-enUS-automotive-medium-2.3.15-pnc.snsr ERROR: With live audio LVCSR and STT models require a VAD. You can add one with the -a flag. % snsr-eval -alt model/stt-enUS-automotive-medium-2.3.15-pnc.snsr NLU intent: no_command (0.9995) = the quick brown fox jumped over the lazy dog's back The quick Brown Fox jumped over the lazy dog's back. ^C ``` Create a new model that includes a VAD: ```console % snsr-edit -vvo vad-stt.snsr \ -t model/tpl-vad-lvcsr-3.17.0.snsr \ -f 0 model/stt-enUS-automotive-medium-2.3.15-pnc.snsr Loading "model/tpl-vad-lvcsr-3.17.0.snsr" as the template model. Loading "model/stt-enUS-automotive-medium-2.3.15-pnc.snsr" into setting "0". Output written to "vad-stt.snsr". % snsr-eval -lt vad-stt.snsr NLU intent: no_command (0.9995) = the quick brown fox jumped over the lazy dog's back The quick Brown Fox jumped over the lazy dog's back. ``` 3. Gate an STT model with a wake word using [tpl-opt-spot-vad-lvcsr](https://doc.sensory.com/tnl/7.8/models/tpl/tpl-opt-spot-vad-lvcsr.md#tpl-opt-spot-vad-lvcsr-type): ```console % snsr-edit -vvo vg-vad-stt.snsr \ -t model/tpl-opt-spot-vad-lvcsr-1.28.0.snsr \ -f 0 model/spot-voicegenie-enUS-6.5.1-m.snsr \ -f 1 model/stt-enUS-automotive-medium-2.3.15-pnc.snsr \ -s include-wake-word-audio=1 Loading "model/tpl-opt-spot-vad-lvcsr-1.28.0.snsr" as the template model. Loading "model/spot-voicegenie-enUS-6.5.1-m.snsr" into setting "0". Loading "model/stt-enUS-automotive-medium-2.3.15-pnc.snsr" into setting "1". Output written to "vg-vad-stt.snsr". ``` [include-wake-word-audio](https://doc.sensory.com/tnl/7.8/api/setting-keys/configuration.md#include-wake-word-audio) `= 1` includes the wake word in the audio seen by the STT recognizer, but configures the STT to elide this from the recognition hypothesis. This improves recognition accuracy if there's no pause between the wake word and the STT command. **Test with snsr-eval** ```console % snsr-eval -lt vg-vad-stt.snsr \ data/audio/voice-genie-set-cruise-control.wav NLU intent: set_cruise_control (0.9968) = set the cruise control to 55 miles per hour NLU entity: number (0.9937) = 55 NLU entity: speed_unit (0.9936) = miles per hour Set the cruise control to fifty five miles per hour. ``` **Test with snsr-eval and live audio** ```console % snsr-eval -t vg-vad-stt.snsr P 7180 7340 Said P 7220 7860 Said the radio P 7220 8220 Set the radio tonight P 7260 8580 Set the radio to ninety one P 7260 9060 Set the radio to ninety one point. Three P 7260 9260 Set the radio to ninety one point. Five P 7260 9620 Set the radio to ninety one point. Five f NLU intent: set_radio (0.9674) = set the radio to 91.5 FM NLU entity: radio_station (0.9688) = 91.5 FM 7260 9860 Set the radio to ninety one point. Five F. M. ^C ``` ## LVCSR _(TrulyNatural only)_ Use [Grammar-based recognition](https://doc.sensory.com/tnl/7.8/models/types/lvcsr.md#grammar-based-recognition) for command and control tasks with small to medium sized vocabularies on devices that aren't powerful enough for STT. [VoiceHub](https://doc.sensory.com/tnl/7.8/reference/voicehub.md#voicehub) provides a convenient interface for creating these models. 1. Build a model that recognizes the example audio files in _data/enrollments/_, using a grammar file from _data/grammars/_: ```console % snsr-edit -vvo commands.snsr \ -t model/lvcsr-build-enUS-14.0.2-5MB.snsr \ -f grammar-stream data/grammars/enrollments-nlu-slot.txt \ -s partial-result-interval=0 Loading "model/lvcsr-build-enUS-14.0.2-5MB.snsr" as the template model. Loading "data/grammars/enrollments-nlu-slot.txt" into setting "grammar-stream". Output written to "commands.snsr". % snsr-eval -t commands.snsr data/enrollments/armadillo-1-0-c.wav NLU intent: calculate (0) = 18 percent of 643 NLU entity: percent (0) = 18 NLU entity: number (0) = 643 375 3150 armadillo 18 percent of 643 ``` [grammar-stream](https://doc.sensory.com/tnl/7.8/api/setting-keys/runtime.md#grammar-stream) is set to the contents of _data/grammars/enrollments-nlu-slot.txt_ to define which sentences the recognizer will accept, and [partial-result-interval](https://doc.sensory.com/tnl/7.8/api/setting-keys/configuration.md#partial-result-interval) `= 0` to suppress interim results. ### Details: _data/grammars/enrollments-nlu-slot.txt_ type: info ``` # LVCSR grammar specification for test utterances in data/enrollments/ # Includes lightweight NLU slot markup. # # In a tpl-spot-vad-lvcsr pipeline the prefix would be consumed by the spotter. prefix = armadillo | jackalope | terminator; # Numbers used in the intent rule below number = 18 | 643 | 20 | 6; # Places place = target | winco | susan's house | gas; # Dates date = friday | tomorrow | next week; # List of known utterances in the *-c.wav files. intent = {calculate {percent $number} percent of {number}} | {call call the nearest {place}} | {navigate how far away is {place}} | {avcontrol {action play} {type more} songs by this artist} | {avcontrol {action record} a {type video}} | {startTimer start a timer for {number} minutes {unit :minutes}} | {navigate i'm running low on {place}} | {calendar {action cancel} all my {type meetings} on {date}} | {navigate directions to {place}} | {messaging do i have any new texts {action :query} {type :texts}} | {calendar {type open} my calendar to {date}} | {alarm {type set} an alarm for {number} am {date}}; # Match the prefix and zero or one of the sentences. # and are sentence start and end markers that # match silence and small amounts of extraneous speech. g = $prefix $intent? ; ``` 2. Combine the wake word [previously enrolled](https://doc.sensory.com/tnl/7.8/getting-started/index.md#qs-ww-enrollment) with this LVCSR model and [tpl-opt-spot-vad-lvcsr](https://doc.sensory.com/tnl/7.8/models/tpl/tpl-opt-spot-vad-lvcsr.md#tpl-opt-spot-vad-lvcsr-type): ```console % snsr-edit -vvo ww-commands.snsr \ -t model/tpl-opt-spot-vad-lvcsr-1.28.0.snsr \ -f 0 udt-kws.snsr \ -f 1 commands.snsr \ -s include-wake-word-audio=1 Loading "model/tpl-opt-spot-vad-lvcsr-1.28.0.snsr" as the template model. Loading "udt-kws.snsr" into setting "0". Loading "commands.snsr" into setting "1". Output written to "ww-commands.snsr". ``` And then run it over all the enrollment recordings: ```console % snsr-eval -t ww-commands.snsr data/enrollments/*-c.wav NLU intent: calculate (0) = 18 percent of 643 NLU entity: percent (0) = 18 NLU entity: number (0) = 643 375 3195 armadillo 18 percent of 643 NLU intent: call (0) = call the nearest target NLU entity: place (0) = target 4695 6360 armadillo call the nearest target NLU intent: navigate (0) = how far away is winco NLU entity: place (0) = winco 7680 9495 armadillo how far away is winco NLU intent: avcontrol (0) = record a video NLU entity: action (0) = record NLU entity: type (0) = video 14535 16095 armadillo record a video NLU intent: startTimer (0) = start a timer for 20 minutes minutes NLU entity: number (0) = 20 NLU entity: unit (0) = minutes 17640 19905 armadillo start a timer for 20 minutes minutes NLU intent: navigate (0) = i'm running low on gas NLU entity: place (0) = gas 21060 22935 jackalope i'm running low on gas NLU intent: calendar (0) = cancel all my meetings on friday NLU entity: action (0) = cancel NLU entity: type (0) = meetings NLU entity: date (0) = friday 24315 26655 jackalope cancel all my meetings on friday NLU intent: navigate (0) = directions to susan's house NLU entity: place (0) = susan's house 27915 30195 jackalope directions to susan's house NLU intent: messaging (0) = do i have any new texts query texts NLU entity: action (0) = query NLU entity: type (0) = texts 31500 33525 jackalope do i have any new texts query texts NLU intent: calendar (0) = open my calendar to next week NLU entity: type (0) = open NLU entity: date (0) = next week 34695 36975 jackalope open my calendar to next week NLU intent: alarm (0) = set an alarm for 6 am tomorrow NLU entity: type (0) = set NLU entity: number (0) = 6 NLU entity: date (0) = tomorrow 38160 40665 jackalope set an alarm for 6 am tomorrow ``` **Note:** type: note The doubled "minutes" in the `startTimer` line above is expected: the grammar rule is: ``` {startTimer start a timer for {number} minutes {unit :minutes}} ``` It matches the spoken word "minutes" against the literal token, then the `{unit :minutes}` slot rewrites to a second "minutes" so the intent text and the `unit` entity carry the same canonical value. See the `:` rewrite operator described in the [grammar syntax reference](https://doc.sensory.com/tnl/7.8/models/types/lvcsr.md#grammar-syntax) for details. [RIFF WAV]: https://en.wikipedia.org/wiki/WAV#RIFF_WAVE "Audio file format" [uv]: https://docs.astral.sh/uv/ "uv Python package and project manager" *[API]: Application Programming Interface *[EFT]: Enrolled Fixed Trigger: fixed wake words adapted to a speaker to improve accuracy *[LPCM]: Linear pulse-code modulation *[LVCSR]: Large Vocabulary Continuous Speech Recognition model, feed-forward neural net acoustic model with FST decoder *[NLU]: Natural Language Understanding model *[PNC]: Punctuation and Capitalization, an STT model variant that emits cased text with punctuation *[SDK]: Software Development Kit *[STT]: Speech To Text: transformers with language model and CTC decoding *[TNL]: TrulyNatural, Sensory's large-vocabulary speech recognition technology *[UDT]: User-Defined Trigger: enrolled wake words and command sets *[VAD]: Voice Activity Detector