Audio Search Using Codec Frames for Real-Time Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current audio search systems require significant processing resources for real-time processing and accuracy, with large indexing phases and storage needs, and can produce errors that affect subsequent searches.
Innovation Solution
The system uses frames from an audio signal to look up entries in a lookup table associated with a codec, determining similarities with sound model elements to detect sound bites, words, or phrases, generating events upon detection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If LVCSR is used for indexing phase, then text representation of audio content is achieved, but large processing resources are required and errors limit search accuracy
Solution Approach 1:
The patent extracts only the necessary acoustic features from audio frames (codec parameters, energy, zero-crossing rate) rather than performing full LVCSR processing. This extraction approach obtains sufficient information for sound bite detection while avoiding the heavy computational burden of complete speech recognition, thereby reducing processing resources while maintaining detection accuracy
Solution Approach 2:
The patent segments the audio signal into frames and processes each frame independently using codec-specific lookup tables. This segmentation allows parallel processing of multiple frames with simplified operations, reducing overall processing complexity compared to sequential LVCSR while maintaining comprehensive coverage of the audio content
2Adaptability or versatility
If statistical models with multiple states are used, then phoneme detection capability is improved, but processing complexity and storage requirements increase
Solution Approach 1:
The patent changes the parameters used for phoneme detection from complex statistical model states to simplified acoustic features including codec parameters, frame energy, and zero-crossing rate. These parameter changes maintain the ability to distinguish different phonemes while dramatically reducing processing complexity and storage requirements
Solution Approach 2:
The patent uses lightweight, frame-by-frame acoustic feature analysis instead of heavy statistical models. Each frame is processed independently with simple feature extraction and lookup table comparison, replacing complex long-term statistical modeling with numerous simple, independent analyses that achieve similar detection capability with far less computational overhead
3Speed
If index data is stored for later search, then search speed is improved, but disk space requirements increase significantly
Solution Approach 1:
The patent stores only the essential acoustic features and codec parameters for each frame in the index, rather than storing complete LVCSR transcripts or complex statistical model data. This partial indexing approach provides sufficient information for rapid sound bite detection while keeping the index size manageable and proportional to the audio duration
Solution Approach 2:
Instead of storing processed speech recognition results and searching through text transcripts, the patent inverts the approach by storing raw acoustic features and performing detection directly on these features during search. This inversion eliminates the need for large transcript storage while maintaining fast search capability through efficient feature matching
Data Source
AI summary
To detect events in an audio stream, frames of an audio signal (e.g., frames generated by a codec for a voice call or music stream) are received. Based on information in the frames, an index is used to look up an entry in a table associated with the codec. Each entry in the table indicates a likelihood that a frame matches a sound model element. The likelihood is used in the search for a sound bite, word, and/or phrase in the audio signal. The process of dynamic programming is used to find the combined likelihood for a match of the word, phrase, and/or sound bite to a region of the audio stream. Upon detection of the word, phrase, and/or sound bite in the audio stream, an event is generated, such as, notifying a person or logging the event in a database.