Wake-Word Speaker Verification with Adaptive Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Voice assistants like BIXBY, SIRI, and ALEXA face challenges with user enrollment being a tedious task, leading to degraded voice wake-up performance and increased invalid wake-ups when enrollment is not required.
Innovation Solution
Implementing a system for self-wake word speaker verification that automatically finds speaker vectors from incoming audio, performs embedding vector updates, and uses clustering to reduce invalid wake-ups without enrollment, with a two-stage similarity threshold for verification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If voice enrollment is required, then voice wake-up performance is improved, but user experience deteriorates due to tedious enrollment tasks
Solution Approach 1:
The system performs automatic speaker verification without requiring users to manually enroll their voices. The voice assistant autonomously captures audio data, extracts embedding vectors, performs clustering to identify speakers, and updates speaker vectors automatically, eliminating the need for user participation in the enrollment process while maintaining accurate voice recognition
Solution Approach 2:
The system pre-processes audio data by extracting embedding vectors and performing clustering in advance to identify speakers before the actual wake-up detection. This preliminary action of creating speaker profiles from ambient audio enables accurate verification without requiring users to undergo explicit enrollment procedures
2Ease of operation
If voice enrollment is eliminated, then user experience is improved, but invalid wake-ups increase
Solution Approach 1:
The system replaces the mechanical enrollment process with an automated machine learning-based verification system. Instead of relying on manual user registration, the system uses embedding vector extraction, clustering algorithms, and speaker verification models to automatically distinguish between valid and invalid wake-up attempts, reducing false positives while maintaining ease of use
Solution Approach 2:
The system continuously updates speaker vectors based on embedded audio data and clustering results. This feedback mechanism allows the system to learn from ambient audio and refine its speaker identification over time, improving wake-up accuracy by adapting to the user's voice characteristics without requiring explicit enrollment
3Ease of operation
If speaker verification is performed without enrollment, then ease of operation is improved, but measurement precision of speaker identification deteriorates
Solution Approach 1:
The system transforms audio data into embedding vectors that capture speaker characteristics in a compressed representation space. By changing the parameter space from raw audio signals to extracted features, the system achieves accurate speaker identification without enrollment. The clustering algorithm then operates on these transformed parameters to group similar voice patterns, enabling precise identification
Solution Approach 2:
The system focuses on extracting and analyzing local characteristics of the voice signal that are distinctive to each speaker. By identifying and emphasizing these local features (such as voice tone, rhythm, and spectral characteristics) within the embedding vectors, the system achieves high speaker identification accuracy even without comprehensive enrollment data
Data Source
AI summary
A method includes obtaining audio data and identifying an utterance of a wake word or phrase in the audio data. The method also includes generating an embedding vector based on the utterance from the audio data and accessing a set of previously-generated vectors representing previous utterances of the wake word or phrase. The method further includes performing clustering on the embedding vector and the set of previously-generated vectors to identify a cluster including the embedding vector, where the identified cluster is associated with a speaker. The method also includes updating a speaker vector associated with the speaker based on the embedding vector and determining, using a speaker verification model, a similarity score between the updated speaker vector and the embedding vector. In addition, the method includes determining, based on the similarity score, whether a speaker providing the utterance matches the speaker associated with the identified cluster.