Version 7 changes¶

TrulyNatural version 7 introduced support for STT and is fully backwards compatible with version 6 models and code.

This library uses semantic versioning.

7.6.1 (2025-11-25)¶

• Changed
  • Android examples build with Java 17 and later.
• Fixed
  • Higher than expected CPU use in wake words and command sets. This addresses a regression introduced in 7.5.0.
  • snsr-debug mistakenly used tpl-opt-spot-vad-lvcsr and stt-enUS-automotive-medium model versions from an earlier SDK release.
  • Some of the C examples failed to build on Windows.
• Security
  • profile could access released memory if called on a Session instance where the model, or a filled template slot, was replaced with another model at runtime.

7.6.0 (2025-09-15)¶

• Added
  • Setting loop = 2 pins the tpl-spot-sequential listening focus to the recognizer in slot 1.
  • tpl-spot-debug-1.5.1.snsr with support for NLU intents and entities.
  • udt-universal-3.67.1.0.snsr with improved accuracy.
  • tpl-spot-vad-3.10.0.snsr with support for include-wake-word-audio.
  • snsr-debug Android sample replaces spot-debug and adds support for STT.
  • tnl tpl-spot-vad-lvcsr-3.20.0.snsr and tpl-opt-spot-vad-lvcsr-1.25.0.snsr with support for include-wake-word-audio.
  • tnl Documentation notes that score values are not usable when result-max > 1.
  • stt stt-enUS-automotive-medium-2.3.15-pnc.snsr
    • Support for nlu-match-max.
    • Setting stt-profile does not change partial-result-interval.
  • stt Support for STT on 64-bit iOS platforms:
    • iPhoneOS on arm64 and arm64e.
    • iPhoneSimulator on arm64 and x86_64.
• Changed
  • Documentation system from Doxygen to Material for MkDocs.
  • arm-linux-gnueabi library requires GLIBC >= 2.17 instead of GLIBC >= 2.34 for improved compatibility.
  • lib/*/README*md for the supported target platforms include the minimum required GLIBC version where relevant.
• Removed
  • tpl-spot-debug-1.5.0.snsr
  • udt-universal-3.66.1.9.snsr
  • tpl-spot-vad-3.9.0.snsr
  • spot-debug Android sample, see snsr-debug instead.
  • Know issues documentation page.
  • tnl tpl-spot-vad-lvcsr-3.19.0.snsr
  • tnl tpl-opt-spot-vad-lvcsr-1.24.snsr
  • stt stt-enUS-automotive-medium-2.3.13-pnc.snsr
• Fixed
  • spot-convert could show incorrect license recognition or duration limits in the comments at the start of the header file it generates.
  • allocPerf did not produce usable results when used with allocStdlib.
  • tnl Incorrect begin-ms, end-ms, and score results for a tpl-spot-vad-lvcsr recognizer in slot 1 of tpl-spot-concurrent.
  • tnl Using tpl-spot-sequential in slot 0 of tpl-opt-spot-vad-lvcsr or tpl-spot-vad-lvcsr could lead to warning messages on stderr.
  • tnl snsr-eval -a option failed when used with class libraries.
  • stt Bus error on 32-bit Arm platforms when using a model with machine-learned NLU or punctuation and capitalization support. This was caused by a bug in ONNX that resulted in an attempt to read int64_t values not aligned to 8 bytes.
• Security
  • tnl If nlu-grammar-stream reduced the length of the phrase result (by using the : rewrite operator, for example) in an STT model with punctuation and capitalization support, an unterminated string could cause an out-of-bounds memory read.

7.5.0 (2025-06-12)¶

• Added
  • snsr-eval -a, -q, and -u flags.
  • stt-support, thread-support.
  • Support for loading parts of a model that resides in ROM onto the heap. This can improve inference speed when ROM is significantly slower than RAM.
  • Reduced peak RAM requirement on small platforms
    • prune:enable option to remove unused settings after a model's been loaded.
    • Reduced overhead in the session configuration store.
  • fromProvider as an alias for snsrStream_alloc().
  • profile is available in the Java language binding.
  • Support for armv7k and arm64_32 watchOS architectures.
  • tpl-spot-concurrent-1.5.0.snsr with support for ^listen-begin, ^listen-end, ^begin, ^end, ^limit, and ^silence.
  • tpl-spot-vad-3.9.0.snsr
  • tnl ram-limit, ac-prune-top-k.
  • tnl Support for improved out-of-grammar detection (with better discrimination and lower RAM overhead) in grammar-based recognition and VoiceHub Pro models.
  • tnl Support for grammar-based recognition models that report recognition scores.
  • tnl Templates with support for domain, am-size, lm-size, nlu-size, and slm-size:
    • tpl-opt-spot-vad-lvcsr-1.24.0.snsr
    • tpl-spot-vad-lvcsr-3.19.0.snsr
    • tpl-vad-lvcsr-3.17.0.snsr
  • tnl Models updated to support score setting:
    • lvcsr-background-enUS-1.2.3.snsr
    • lvcsr-build-enUS-2.7.3.snsr
  • stt domain, slm-size.
  • stt stt-enUS-automotive-medium-2.3.13-pnc.snsr with support for domain and slm-size. This model includes a single language model, so domain will return NULL. It does not include a generative language model either: slm-size is 0.
• Changed
  • Documentation available online.
  • snsr-edit requires the -o flag to create output.
  • snsr-eval prints scores with the %g format specifier.
  • getDouble, setDouble, getInt, and setInt support both double and integer settings, but will report INCORRECT_SETTING_TYPE if a double value cannot be converted to an integer without loss of precision or overflow.
  • getString and setString support double and integer settings.
  • set no longer requires quotes for string arguments without embedded spaces.
  • SOURCE, SOURCE_RAM, and SOURCE_PRUNED create code that tags all model objects with preprocessor macro SNSR_MODEL_ATTR. Use this, for example, to specify the section in which the linker places these objects.
  • sample/c/CMakeLists.txt
    • Builds snsr-eval-subset.
    • Installs the sample binaries in sample/c/bin/.
  • sample/c/Makefile and sample/c/CMakeLists.txt build the live-spot-stream.c example on macOS.
  • util/gradlew uses Gradle 8.14, compatible with Java 21.
  • Java examples use lambda expressions for event handlers and iterators.
  • Enrolled wake word tasks no longer invoke ^pause and ^resume when interactive == 0.
  • tnl tpl-opt-spot-vad-lvcsr and tpl-spot-vad-lvcsr with include-leading-silence = 1 no longer invoke the ^result, ^nlu-intent, and ^nlu-slot events if the VAD does not detect speech. The ^listen-end directly follows the ^silence event.
  • tnl snsr-eval with live audio input will not run LVCSR or STT models that do not include a VAD component. Use the new -a flag to add tpl-vad-lvcsr on the fly.
  • tnl complete-only = 1 accepts results even if there's no trailing silence in the utterance.
  • stt In STT models with punctuation and capitalization support, empty results are reported as an empty string instead of a single period.
• Removed
  • tpl-spot-concurrent-1.4.0.snsr
  • tpl-spot-vad-3.8.0.snsr
  • tnl lvcsr-background-enUS-1.2.1.snsr
  • tnl lvcsr-build-enUS-2.7.0.snsr
  • tnl tpl-opt-spot-vad-lvcsr-1.20.0.snsr
  • tnl tpl-spot-vad-lvcsr-3.11.0.snsr
  • tnl tpl-vad-lvcsr-3.10.0.snsr
  • stt stt-enUS-automotive-medium-2.3.13.snsr
• Fixed
  • Machine-learned VAD components detected the speech onset late on synthetic audio data with significant all-zero leading silence.
  • oss-components always returned NULL.
  • tnl In rare cases grammar-based recognition NLU intent names could mistakenly include the first word of the NLU intent value.
  • tnl In rare cases VoiceHub models that use @ref build could misclassify in-domain intents as background.
  • tnl Potentially reduced recognition accuracy in grammar-based recognition models that use ram-limit to constrain the Viterbi decoder.
• Security
  • Deeply-nested templates with duplicate models (for example, combining the same spotter model with tpl-spot-concurrent, then combining the result with another tpl-spot-concurrent and the same spotter model, three levels deep), could result in out-of-bounds memory access when releasing the session.

7.4.1 (2025-03-13)¶

• Fixed
  • stt Using custom-vocab to map multi-word sequences could result in missed NLU entities.

7.4.0 (2025-01-22)¶

• Added
  • Java support on x86_64-windows-msvc.
  • SOURCE_RAM and snsr-edit -C option.
  • Spotter template modes with new settings low-fr-operating-point and duration-ms:
    • tpl-spot-debug-1.5.0.snsr
    • tpl-spot-vad-3.8.0.snsr
    • tpl-spot-sequential-1.5.0.snsr
  • Support for Smart Wake Words. On a near-miss false reject, these models switch to a more accepting operating point for a limited duration to make it more likely that the model will accept a repeated wake word utterance. Enable this feature (on wake word models that support it) by setting low-fr-operating-point. Adjust the duration of the low false-reject listening window with duration-ms.
  • spot-voicegenie-enUS-6.5.1-m.snsr with support for low-fr-operating-point and duration-ms.
  • Compatibility with 2025 VoiceHub Pro wake words and command sets.
  • VERSION_CMAKE.
  • tnl tpl-opt-spot-vad-lvcsr-1.20.0.snsr supports changing VAD settings backoff, hold-over, leading-silence, max-recording, trailing-silence, and include-leading-silence. Adds vocab-iterator, ^slm-start, ^slm-result-partial, ^slm-result, slm-enabled, slm-turn-limit, stt-profile, low-fr-operating-point, and duration-ms.
  • tnl VAD template models with new settings ^slm-start, ^slm-result-partial, ^slm-result, slm-enabled, slm-turn-limit, stt-profile, low-fr-operating-point, and duration-ms:
    • tpl-spot-vad-lvcsr-3.15.0.snsr
    • tpl-vad-lvcsr-3.14.0.snsr
  • stt stt-profile.
  • stt Experimental support for generative language models. New settings: ^slm-start, ^slm-result-partial, ^slm-result, slm-enabled, and slm-turn-limit.
  • stt stt-enUS-automotive-medium-2.3.13.snsr
    • Results include capitalization and punctuation.
    • Supports stt-profile.
• Changed
  • Android JNI shared libraries export only TrulyNatural function names. This significantly reduces the chances of symbol clashes in applications that use third-party libraries that depend on the same OSS modules as the TrulyNatural SDK.
  • Android JNI libraries built with NDK r27c and flexible page size support.
  • Improved compatibility with older Android releases, support for API level 21 and later.
  • SnsrLibrary.cmake removes unused functions and data from target executables (-ffunction-sections -fdata-sections)
  • DSP library (for creating embedded spotter models) version 8.5.0.
  • Toolchain updates:
    • arm-linux-gnueabi built with gcc 13.2.0.
    • arm-linux-gnueabihf built with gcc 13.3.1.
    • i686-linux-gnu built with gcc 13.2.0.
  • On Windows, snsr-eval sets the console code page to UTF-8.
  • If application code used push and did not call stop before releasing the model, the call to release invoked all pending event handlers. It now ignores these.
  • stop completes even if the session is in an error state.
  • tnl No ^result-partial for empty hypotheses.
  • stt Support Windows \r\n line termination as entry separators in custom-vocab.
  • stt Removed the 1000-entry limit for custom-vocab. Please note that using such large custom vocabularies is frequently counter-productive. Consider contacting your Sensory Sales representative to discuss creating a domain-specific recognizer instead.
  • stt Improved custom-vocab behavior in character-based languages such as Mandarin Chinese.
  • stt STT models start fewer threads.
• Deprecated
  • threshold.
• Removed
  • spot-voicegenie-enUS-6.5.0-m.snsr
  • tpl-spot-debug-1.4.0.snsr
  • tpl-spot-sequential-1.4.0.snsr
  • tpl-spot-vad-3.7.0.snsr
  • tpl-spot-dynop-1.4.0.snsr
  • tnl tpl-opt-spot-vad-lvcsr-1.12.0.snsr
  • tnl tpl-spot-vad-lvcsr-3.11.0.snsr
  • tnl tpl-vad-lvcsr-3.10.0.snsr
  • stt stt-enUS-automotive-small_medium-2.2.13-BBB-ff.snsr
• Fixed
  • With push-duration-limit > 0, reset did not flush the push backlog buffer enabled with push-buffer-size.
  • Low memory use VoiceHub Large Vocabulary Project models could fail to run if settings (such as operating-point) were changed after registering event handlers.
  • Reading settings from models built from deeply nested templates could mistakenly report VALUE_NOT_SET.
  • tnl search.frame-nota was ignored for ^result when used with lightweight NLU.
• Security
  • If application code using the C API used push, and a result handler called forEach, and that handler was invoked by release because the session was released without calling stop first, then a double-free and access of invalid heap memory could occur.
  • tnl Out-of-bounds memory access could occur if tpl-opt-spot-vad-lvcsr was used with 0.include-leading-silence = 1.

7.3.0 (2024-07-29)¶

• Added
  • New FAQ section: How do I take action on an NLU result?
  • Support for adding this SDK to projects that use CMake. See C examples.
  • Real-time clock implementation for Windows, enables profile:real-time.
  • Support for multi-phrase spotter models with lower false accept rates.
  • arm-none-eabi library with NEON built with gcc 13.3.1.
  • tnl VAD template models with settings backoff and max-recording optimized for STT. The defaults are now 300 ms for audio backoff, and 60 seconds for the maximum recording duration.
    • tpl-opt-spot-vad-lvcsr-1.12.0.snsr
    • tpl-spot-vad-lvcsr-3.11.0.snsr
    • tpl-vad-lvcsr-3.10.0.snsr
  • stt Improved STT recognition speed and accuracy, and reduced ^result and ^nlu-intent latency when used with a VAD.
  • stt STT support for arm-linux-gnueabihf, aarch64-linux-gnu, and x86_64-windows-msvc.
  • stt stt-enUS-automotive-small_medium-2.2.13-BBB-ff.snsr
    • 100 MiB total size compared to 299 MiB in previous version.
    • Improved NLU inference speed.
    • Improved accuracy.
• Changed
  • arm-linux-gnueabihf and aarch64-linux-gnu libraries built with gcc 10.3.1.
  • tnl Optimized machine-learned NLU models. On average these now run about ten times faster.
• Removed
  • Windows ports: x86-windows-2008, x86-windows-2015, x86_64-windows-2008, x86_64-windows-2015, x86_64-windows-mingw.
  • arm-buildroot-linux-uclibcgnueabihf library.
  • arm-ca7-linux-gnueabihf library.
  • tnl VAD template models:
    • tpl-opt-spot-vad-lvcsr-1.11.0.snsr
    • tpl-spot-vad-lvcsr-3.10.0.snsr
    • tpl-vad-lvcsr-3.9.0.snsr
  • stt stt-enUS-automotive-small_medium-2.0.8-BBB-ff.snsr
• Fixed
  • Certain low false-accept spotter models reported an internal validation error when used with tpl-spot-sequential-1.4.0.snsr
  • spot-convert -c produced incorrect code output for recent low false-accept models.
  • The iOS sample app mistakenly embedded snsr.xcframework, leading to TestFlight rejection due to asset validation errors.
  • The TrulyNatural SDK linked against -landroid, making it incompatible with the Vendor Native Development Kit.
  • stt Duplicate entities and typographic errors in the documentation for stt-enUS-automotive-small_medium-2.2.13-BBB-ff.snsr
• Security
  • stt NLU models could cause out-of-bounds memory access on very long (hundreds of words) recognition results.

7.2.0 (2024-04-11)¶

• Added
  • snsr-eval -p option shows an estimate of the CPU processing capacity required to run a model in real-time.
  • profile:real-time.
  • tnl nlu-intent-score, nlu-entity-score, and nlu-slot-score.
  • tnl Models updated to support NLU score and score settings:
    • tpl-opt-spot-vad-lvcsr-1.11.0.snsr
    • tpl-spot-vad-lvcsr-3.10.0.snsr
    • tpl-vad-lvcsr-3.9.0.snsr
  • stt am-size, lm-size, nlu-size.
  • stt stt-enUS-automotive-small_medium-2.0.8-BBB-ff.snsr
    • Improved recognition accuracy compared to stt-v2-enUS-automotive-0.2.11-BT-400ms.snsr
    • Adjusts model evaluation interval to match partial-result-interval.
• Changed
  • run ignores push-duration-limit.
  • tnl snsr-eval -lll does not show ^nlu-intent or ^nlu-slot results.
  • tnl snsr-eval shows NLU intent, entity, and slot scores.
  • stt STT models use partial-result-interval to set the model evaluation interval. The larger this interval the lower the average CPU use.
• Removed
  • tnl Models updated to support NLU score settings:
    • tpl-opt-spot-vad-lvcsr-1.9.0.snsr
    • tpl-spot-vad-lvcsr-3.8.0.snsr
    • tpl-vad-lvcsr-3.8.0.snsr
    • tpl-spot-lvcsr-1.5.0.snsr (Use tpl-spot-vad-lvcsr-3.10.0.snsr instead.)
  • stt stt-v2-enUS-automotive-0.2.11-BT-400ms.snsr
• Fixed
  • load now returns LIBRARY_TOO_OLD when attempting to load a model that uses features it does not support.

7.1.1 (2024-02-13)¶

• Fixed
  • Linking against the Android static libraries could suppress the output of calls to __android_log_print(), __android_log_vprint(), and __android_log_write() in the entire application.
  • The AAR file for Android once again links against -llog.
  • tnl tpl-spot-vad-lvcsr-3.8.0.snsr would keep listening for LVCSR audio if enough silence after the wake word lead to ^silence.
    • This happened only for LVCSR recognizers (including those made with VoiceHub) and did not affect STT ones.
    • Regression from 7.0.0.

7.1.0 (2024-01-11)¶

• Added
  • tnl Simplified NLU API
    • Reduces NLU slot hierarchy to just two levels.
    • Provides direct access to top-level NLU slots as "intents" and child slots as "entities".
    • Does not support nlu-match-max > 1.
    • Does not replace the existing NLU slot API.
    • See ^nlu-intent, nlu-intent-name, nlu-intent-value, nlu-entity-iterator, nlu-entity-count, nlu-entity-name, and nlu-entity-value.
  • tnl Models updated to support simplified NLU API:
    • lvcsr-background-enUS-1.2.1.snsr
    • lvcsr-build-enUS-2.7.0.snsr
    • tpl-opt-spot-vad-lvcsr-1.9.0.snsr
    • tpl-spot-lvcsr-1.5.0.snsr
    • tpl-spot-vad-lvcsr-3.8.0.snsr
  • stt Support for machine-learned NLU intent and entity classifiers.
  • stt custom-vocab
  • stt stt-v2-enUS-automotive-0.2.11-BT-400ms.snsr
• Changed
  • AAR file for Android is no longer linked against -landroid and -llog. This allows it to be used in an AOSP system image.
• Removed
  • Energy-only VAD models superseded by machine-learned versions:
    • tpl-spot-vad-2.4.0.snsr
    • tpl-spot-vad-2.0.1.snsr
    • tnl tpl-spot-vad-lvcsr-2.4.0.snsr
    • tnl tpl-vad-lvcsr-2.5.0.snsr
  • tnl Models updated to support simplified NLU API:
    • lvcsr-background-enUS-1.1.0.snsr
    • lvcsr-build-enUS-2.6.0.snsr
    • tpl-opt-spot-vad-lvcsr-1.8.0.snsr
    • tpl-spot-lvcsr-1.4.0.snsr
    • tpl-spot-vad-lvcsr-3.7.0.snsr
    • tpl-vad-lvcsr-3.7.0.snsr
  • stt stt-v1-en-1.0.1-400ms.snsr
• Fixed
  • VAD models models ignored the max-recording setting.
  • tnl tpl-opt-spot-vad-lvcsr did not produce a ^result for slot = 0 if 0.include-leading-silence was set to 1 and a leading-silence timeout occurred.
  • tnl LVCSR models used in a VAD template would stop producing ^result callbacks if the VAD reported ^limit. STT models were not affected.

7.0.0 (2023-11-20)¶

• Added
  • CONFIG_PRUNED and SOURCE_PRUNED reduce the size of serialized models by removing configuration options such as unused operating-point values.
  • snsr-edit -p option reduces phrasespot model sizes by removing operating points other than the currently selected operating-point.
  • stt STT support with improved accuracy suitable for command-and-control and dictation tasks.
    • This initial release includes STT support on Android, macOS, and x86_64-linux only.
    • Includes Open Source modules, see open source licenses.
    • Requires linking against a C++ runtime, see Integrate with your build.
    • Using STT models do not require any API changes. These are drop-in replacements for older LVCSR models on supported platforms.
    • STT recognition models
      • stt-v1-en-1.0.1-400ms.snsr
      • Contact your Sensory Sales representative for additional languages.
    • New settings: oss-components, STT_SUPPORT.
• Changed
  • The deprecated delay setting is read-only.
  • DSP library (for creating embedded spotter models) version 7.4.0.
  • tnl The effects of : terminal rewrites @b are now visible in ^result text values.
  • tnl The lightweight NLU processor scoring algorithm no longer considers :inserted and removed: words when ranking result parses. This potentially changes the primary NLU result, or the order in which NLU results are presented when nlu-match-max > 1, for grammars that use the : rewrite feature to better match expectations.
  • tnl If the time required to compute a partial LVCSR result exceeds the push-duration-limit, the ^result-partial callback does not happen.
  • tnl The grammar-based recognition grammars support the literal use of special tokens by escaping them with a backslash.
• Deprecated
  • Lightweight VAD. Use a VAD model instead.
• Removed
  • tnl References to the <- slot assignment operator in grammar-based recognition. This had been a symbol reserved for internal use only.
  • tnl lvcsr-broad-enUS-3.1.0-t5.snsr
• Fixed
  • Machine-learned VAD models could unexpectedly report an error that from-index (...) must be >= first-index (...)
  • A subset of spotter models showed increased false accept rates when used with task templates.
  • tnl tpl-opt-spot-vad-lvcsr mistakenly invoked ^silence and ^listen-end multiple times at end-of-stream if 0.include-leading-silence was set to 1.
  • tnl tpl-spot-concurrent with a spotter model in slot 0 and an LVCSR model (gated by a VAD, or spotter and VAD) in slot 1 would mistakenly return nlu-slot-name and nlu-slot-value values as NULL.
  • tnl ^nlu-slot could still occur even if complete-only suppressed the recognition result.
  • tnl grammar-based recognition with classes could fail with a segmentation fault when attempting to build a recognizer with too many class references.
• Security
  • This sequence of operations could lead to getString returning an unterminated string:
    • Use setStream to set a configuration value to a sequence of bytes that does not include \0.
    • Save the model to persistent storage.
    • Load the modified model from persistent storage.
    • Access the configuration value with getString.
  • In the Java language binding fromMemory and fromString could potentially access the underlying byte[] memory after it was garbage-collected.