Abstract: Systems and methods for generating sound profiles of artificial commands detected by multiple voice activated electronic devices is described herein. In some embodiments, numerous voice activated electronic devices may send audio data representing a phrase to a backend system at a substantially same time. Text data representing the phrase, and counts for instances of that text data, may be generated. If the number of counts exceeds a predefined threshold, the backend system may cause any remaining response generation functionality that particular command that is in excess of the predefined threshold to be stopped, and those devices returned to a sleep state. In some embodiments, a sound profile unique to the phrase that caused the excess of the predefined threshold may be generated such that future instances of the same phrase may be recognized prior to text data being generated, conserving the backend system's resources.

Abstract: User-specific language models (LMs) that include internal word indexes to a word table specific to the user-specific LM rather than a word table specific to a system-wide LM. When the system-wide LM is updated, the word table of the user-specific LM may be updated to translate the user-specific indices to system-wide indices. This prevents having to update the internal indices of the user-specific LM every time the system-wide LM is updated.

Abstract: Systems and methods for detecting similar audio being received by separate voice activated electronic devices, and ignoring those commands, is described herein. In some embodiments, a voice activated electronic device may be activated by a wakeword that is output by the additional electronic device, such as a television or radio, may capture audio of sound subsequently following the wakeword, and may send audio data representing the sound to a backend system. Upon receipt, the backend system may, in parallel to performing automated speech recognition processing to the audio data, generate a sound profile of the audio data, and may compare that sound profile to sound profiles of recently received audio data and/or flagged sound profiles. If the generated sound profile is determined to match another sound profiles, then the automated speech recognition processing may be stopped, and the voice activated electronic device may be instructed to return to a keyword spotting mode.

Abstract: In a dynamic automatic speech recognition (ASR) processing system, ASR processing may be configured to estimate a latency of returning speech results to a user based on work being done by an ASR processor. The ASR processing system may measure work done by an ASR processor by measuring one or more time independent metrics and comparing the metrics to threshold values. If the metrics exceed the thresholds, the ASR system may take steps to reduce latency associated with processing the utterance, including adjusting a speech recognition parameter.

Abstract: In an automatic speech recognition (ASR) processing system, ASR processing may be configured to reduce a latency of returning speech results to a user. The latency may be determined by comparing a time stamp of an utterance in process to a current time. Latency may also be estimated based on an endpoint of the utterance or other considerations such as how difficult the utterance may be to process. To improve latency the ASR system may be configured to adjust various processing parameters, such as graph pruning factors, path weights, ASR models, etc. Latency checks and corrections may occur dynamically for a particular utterance while it is being processed, thus allowing the ASR system to adjust to rapidly changing latency conditions.

Abstract: Determining the end of an utterance for purposes of automatic speech recognition (ASR) may be improved with a system that provides early results and/or incorporates semantic tagging. Early ASR results of an incoming utterance may be prepared based at least in part on an estimated endpoint and processed by a natural language understanding (NLU) process while final results, based at least in part on a final endpoint, are determined. If the early results match the final results, the early NLU results are already prepared for early execution. The endpoint may also be determined based at least in part on the content of the utterance, as represented by semantic tagging output from ASR processing. If the tagging indicate completion of a logical statement, an endpoint may be declared, or a threshold for silent frames prior to declaring an endpoint may be adjusted.

Abstract: Features are disclosed for identifying and providing command suggestions during automatic speech recognition. As utterances are interpreted, suggestions may be provided based on even partial interpretations to guide users of a client device to commands available via speech recognition.