Voice Command Recognition Using Segmented Language Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems often misinterpret voice commands due to generic language models not trained on individual names, leading to incorrect transcriptions and low recognition confidence scores, especially when users provide commands like "Call Barack Obama".
Innovation Solution
A method involving a mobile computing device that transmits audio data to a server, which applies a first language model for initial transcription and then uses a second, specific language model trained on user contact names to accurately identify the intended contact, allowing the system to perform actions like initiating a phone call.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a generic language model is used for speech recognition, then the system can process various languages and dialects, but the accuracy of transcribing specific contact names deteriorates
Solution Approach 1:
The language model is segmented into two distinct components: a first language model for general language processing and a second language model specifically trained on contact names. This segmentation allows each model to specialize in its respective function, resolving the contradiction between general adaptability and specific transcription accuracy.
Solution Approach 2:
The second language model is trained with a limited vocabulary specifically tailored to the user's contact names, creating local quality specialization. This localized training enables high accuracy for contact name recognition while the first language model maintains general language understanding capabilities.
2Measurement precision
If a second language model trained on contact names is applied, then the accuracy of identifying contact names improves, but the complexity of the speech recognition system increases
Solution Approach 1:
The system is segmented into distinct processing stages with separate language models for different functions. The first language model handles general transcription while the second handles contact name identification, allowing complexity to be organized and managed through functional separation rather than monolithic design.
Solution Approach 2:
The server computing system acts as an intermediary that receives audio data, applies the first language model, determines communication-related voice commands, and then provides transcriptions to the mobile device for further processing with the second language model. This intermediary architecture distributes complexity across multiple components.
3Speed
If voice commands are processed locally on the mobile device, then the response time is reduced, but the computational resources required increase
Solution Approach 1:
Computational tasks are segmented between the mobile device and server. The mobile device performs audio capture and initial processing with the first language model, while complex contact name identification using the second language model is performed on the server. This segmentation optimizes the balance between response time and computational resource consumption.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for obtaining audio data corresponding to an utterance; transmitting the audio data corresponding to the utterance; receiving an indication that that utterance likely includes a communication-related voice command; in response to receiving the indication that the utterance likely includes the communication-related voice command, applying at least a language model to a representation of the audio data corresponding to the utterance, to identify data referencing a contact; and transmitting the data referencing the contact.