Reverse ASR for Pre-Wakeword Command Recognition

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Conventional voice virtual assistants struggle to recognize commands spoken before a wakeword efficiently, leading to issues like capturing and processing pre-command speech inaccurately and consuming excessive power due to continuous voice activity detection.

Innovation Solution

Implementing reverse automatic speech recognition (R-ASR) to process speech audio in the reverse order, starting from the wakeword, to distinguish and process command speech from non-command speech, reducing the need for continuous voice activity detection and minimizing power consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional voice virtual assistants use continuous voice activity detection to process pre-wakeword speech, then command recognition capability is improved, but power consumption increases excessively

Engineering Contradiction:

Improvecommand recognition capabilityVSAvoidpower consumption

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The patent applies reverse automatic speech recognition that processes audio in reverse order (from wakeword backward to command) instead of conventional forward processing. This inversion allows the system to identify commands without continuous voice activity detection, significantly reducing power consumption while maintaining recognition accuracy.

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The system performs preliminary wakeword detection and buffers audio in advance, then processes only the relevant portions using reverse ASR. This preliminary action eliminates the need for continuous voice activity detection during the entire audio stream, reducing power consumption while ensuring command recognition.

Inventive Principle:

Principle #10Preliminary action

2Adaptability or versatility

If conventional virtual assistants buffer a fixed number of seconds of audio before the wakeword, then pre-wakeword speech processing is enabled, but inaccurate capture and processing of pre-command speech occurs

Engineering Contradiction:

Improvepre-wakeword speech processing capabilityVSAvoidcommand speech boundary detection accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The system dynamically adjusts the audio buffer size based on the detected command duration and type rather than using a fixed time window. This dynamic approach allows accurate identification of command boundaries even when commands vary in length or are interspersed with non-command speech, eliminating false captures.

Inventive Principle:

Principle #15Dynamics

Solution Approach 2:

The patent replaces the mechanical fixed-time-buffering approach with reverse automatic speech recognition that uses linguistic and acoustic analysis to dynamically identify command boundaries. This substitution enables precise differentiation between command speech and non-command speech regardless of buffer size.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

3Reliability

If conventional assistants scan for silence before commands in buffered audio, then command identification is attempted, but processing delays increase and false matches occur

Engineering Contradiction:

Improvecommand identification accuracyVSAvoidprocessing delay

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

By processing audio in reverse order from the wakeword backward, the system immediately identifies commands without needing to scan through entire buffered segments for silence markers. This reverse processing eliminates scanning delays and prevents false matches caused by silence detection in non-command contexts.

Inventive Principle:

Principle #13The other way round (Inversion)

Solution Approach 2:

The system extracts and processes only the relevant command portions of audio using reverse ASR, rather than scanning entire buffered segments for silence. This extraction approach eliminates unnecessary scanning time and reduces processing delays while maintaining high accuracy.

Inventive Principle:

Principle #2Taking out (Extraction)

Data Source

PatentUS20250391406A1Pre-wakeword speech processing with reverse automatic speech recognition

Publication Date: 2025.12.25 SOUNDHOUND AI IP LLC

AI summary

Methods and systems for pre-wakeword speech processing are disclosed. Speech audio, comprising command speech spoken before a wakeword, may be stored in a buffer in oldest to newest order. Upon detection of the wakeword, reverse acoustic models and language models, such as reverse automatic speech recognition (R-ASR) can be applied to the buffered audio, in newest to oldest order, starting from before the wakeword. The speech is converted into a sequence of words. Natural language grammar models, such as natural language understanding (NLU), can be applied to match the sequence of words to a complete command, the complete command being associated with invoking a computer operation.