ProxiMic: Convenient Voice Activation via Close-to-Mic Speech Detected by a Single Microphone

Title of the Paper

ProxiMic: Convenient Voice Activation via Close-to-Mic Speech Detected by a Single Microphone

Paper Information

• Research Area: Human-Computer Interaction, Voice Activation Technology
• Keywords: Voice Input, Sensing Technology, Activity Recognition, Close-to-Mic Speech, Privacy Protection, User Experience, Machine Learning, Human-Computer Interaction, Perception Technology, Adverse Environment Detection

Research Background and Problem

• Challenges:
  • Voice input faces three main challenges: prolonged wake-up time, cumbersome and frustrating multi-turn interaction processes, and risks of voice privacy leakage.
  • Existing methods such as "Raise to Speak" and "PrivateTalk" have made breakthroughs in reducing wake-up time, but they rely on multiple sensors or specific hardware configurations, limiting their applicability.
  • Traditional privacy protection methods like silent speech are effective but involve complex devices and support limited semantic ranges, preventing widespread adoption.
• Importance of the Problem:
  • With the increasing prevalence of voice input technology, addressing voice privacy and interaction experience issues is crucial for enhancing user experience.
• Research Motivation and Related Work:
  • To provide a wake-free, low-cost voice input method that addresses wake words, cumbersome interactions, and privacy concerns.
  • The proposed technology can adapt to various device formats without requiring additional sensors or specialized equipment.

Solution

• Method or Solution:
  • A technology named ProxiMic is proposed, enabling wake-free activation via detecting "close-to-mic speech" using a single microphone. It supports users in conducting direct voice interactions across multiple conversational turns.
  • CNN (Convolutional Neural Network) is employed to detect "pop noise" and other subtle close-to-mic features.
  • A two-stage algorithm is designed:
    • The first stage uses an Adaptive Amplitude Threshold Trigger (AATT) to filter potential close-to-mic speech signals.
    • The second stage employs CNN to determine whether the signal is close-to-mic speech for further precise identification.
• Innovations:
  • Innovative use of pop noise as a sound feature for voice activation without requiring multi-sensor support.
  • Development and optimization of a two-stage detection algorithm, balancing high accuracy, low power consumption, and low memory usage.
• Implementation Steps and Key Technologies:
  • AATT: Dynamically monitors environmental background sound energy to filter high-amplitude potential signals.
  • CNN: Refines and distinguishes close-to-mic speech signals based on sound spectrogram features (low-frequency enhanced spectrogram).
  • Experiments simulate real-world scenarios with various device environments and validate robustness against complex background noise.

Research Results

• Specific Results:
  • Achieved a 94.1% activation recall rate and 12.3 false accepts per week per user.
  • Demonstrated adaptability of ProxiMic across various device formats (e.g., smartphones, watches, headphones) with high environmental robustness.
• Advantages:
  • Does not require special gestures or additional sensors, making it low-cost and easy to deploy.
  • Significantly enhances voice input privacy, allowing users to issue voice commands with low volume or whispering.
  • Outperforms existing algorithms (e.g., "Raise to Speak") in memory usage and processing efficiency.
• Experimental or Evaluation Results:
  • User experiments indicate that compared to traditional wake words or button activation methods, ProxiMic achieves higher efficiency ratings and user satisfaction.
  • ASR experimental results show high transcription accuracy even with pop noise or low-volume whispering.
• Limitations and Future Directions:
  • Multi-device adaptation requires further data training and testing to support various hardware configurations.
  • Optimizing semantic recognition for whispering and further reducing false trigger rates are key future research directions.
  • Enhancing robustness in more complex linguistic environments and adverse acoustic scenarios remains a challenge.

Conclusion

This study proposes an innovative and efficient close-to-mic voice activation technology, addressing the interaction complexity and privacy issues inherent in traditional wake-up and voice input methods. Experimental results demonstrate that ProxiMic significantly improves user experience and has broad application potential, such as privacy-sensitive scenarios and IoT device voice interactions. The research provides insights for the future development of voice input technology while highlighting challenges and opportunities in multi-device adaptation and feature expansion.