A distributed system for recognizing home automation commands and distress calls in the Italian language

Emanuele Principi, Stefano Squartini, Francesco Piazza, Danilo Fuselli, Maurizio BonifaziInterspeech 2013

TL;DR

This paper introduces a distributed system designed to enhance safety and convenience in smart homes by recognizing Italian home automation commands and, crucially, distress calls. The system leverages a network of local units to monitor the acoustic environment, enabling automatic emergency calls and hands-free communication when distress is detected.

What Problem Does This Paper Address?

As the global population ages, there's a growing need for technologies that support independent living and ensure the safety of elderly individuals in their homes. Traditional assistive technologies often rely on invasive sensors. This paper addresses the challenge of creating a less intrusive system that can reliably detect both everyday home automation commands and urgent distress calls using only audio input. The goal is to provide a seamless way for individuals to control their environment and, more importantly, to automatically alert help in emergencies, bridging the gap between technological assistance and personal safety.

Methods

The proposed system is architected as a distributed network comprising Local Multimedia Control Units (LMCUs) and a Central Management and Processing Unit (CMPU).

  • LMCU: Each LMCU is equipped with microphones to continuously monitor the acoustic environment. It employs a Voice Activity Detector (VAD) to identify speech segments. For speech recognition, it utilizes the PocketSphinx engine, which is optimized for embedded devices. To enhance robustness against noise and reverberation, the system incorporates Power Normalized Cepstral Coefficients (PNCC) for feature extraction and an adaptive algorithm for interference cancellation (e.g., from TVs or radios). The LMCU operates in two states: "monitoring" (where speech recognition is active) and "calling" (when a phone call is in progress, activating Acoustic Echo Cancellation (AEC) and a Voice over IP (VoIP) stack).
  • Acoustic Echo Cancellation (AEC): The Speex codec's AEC algorithm is used to mitigate echoes in hands-free communication, ensuring call quality. This same algorithm is adapted to reduce known interference sources.
  • Speech Recognition: PocketSphinx is used for recognizing both commands and distress calls. The system employs a simple finite state grammar that combines home automation commands and distress calls. To handle out-of-grammar utterances and prevent false positives, a "garbage phone" technique inspired by Vertanen is implemented.
  • Speaker Adaptation: To improve recognition accuracy, especially for individual users and varying acoustic conditions, speaker adaptation is performed using Maximum Likelihood Linear Regression (MLLR) on phonetically rich sentences.

Datasets

  • APASCI Corpus: Used for training the acoustic model. This corpus consists of Italian speech utterances recorded in a quiet room, with a vocabulary of about 3000 words.
  • ITAAL Corpus: A newly introduced corpus specifically created for evaluating the system's performance. It comprises utterances from 20 native Italian speakers (half male, half female) speaking home automation commands, distress calls, and phonetically rich sentences. Crucially, each sentence was recorded in both normal and shouted speaking styles, and using both a headset microphone and a distant microphone array, simulating realistic home environments with varying noise and reverberation levels.

Key Insights and Ideas

  • Distributed Monitoring for Safety: The distributed nature of LMCUs allows for comprehensive acoustic coverage of a home, enabling proactive detection of emergencies without requiring users to be in a specific location.
  • Robustness through PNCC and Interference Cancellation: The use of PNCC features and adaptive interference cancellation significantly improves the system's ability to recognize speech in noisy and reverberant home environments, which is a common challenge for such systems.
  • Dual-Purpose Recognition: The system effectively integrates the recognition of mundane home automation commands with critical distress calls, offering both convenience and a vital safety net.
  • Importance of Speaker Adaptation: Speaker adaptation, even with limited data, demonstrably improves recognition accuracy, especially in challenging acoustic conditions and when dealing with accents different from the training data.
  • Shouted Speech Challenges: Recognizing shouted speech, particularly for commands, presents a significant challenge, highlighting the need for specialized techniques for this vocal effort.

Key Results

  • Headset Microphone Performance: Without speaker adaptation (Baseline), the system achieved over 90% accuracy for commands and distress calls in normal speaking styles using a headset microphone.
  • Distant Microphone Performance: Performance was lower with the distant microphone due to acoustic mismatches, but speaker adaptation (MLLR) significantly improved accuracy, reaching an average of 21.67% improvement.
  • Distress Call Recognition: With speaker adaptation, the system achieved a distress call recognition accuracy of 85.00% in normal speaking styles and 72.67% in shouted speaking styles using the distant microphone. The authors consider this satisfactory for real-world deployment.
  • Shouted Command Recognition: Shouted command recognition showed a notable decrease in accuracy, even with adaptation, indicating this is an area requiring further improvement.
  • CPU Occupancy: The recognition engine (including feature extraction) had a CPU occupancy of approximately 20%.

Limitations and Future Work

  • Shouted Speech Recognition: The paper acknowledges that recognizing shouted speech, especially for commands, still needs improvement. Future work will focus on algorithms to better detect speaking style and improve recognition in these conditions.
  • Distant Talking Conditions: While speaker adaptation improved performance, further enhancements are needed for distant talking scenarios. The authors plan to explore using multiple microphone channels to boost performance.
  • Subjective Evaluation: The current evaluation is primarily quantitative. Future work will include subjective testing to assess the overall user experience and effectiveness of the complete system.
  • Interference Cancellation in Real Scenarios: The interference cancellation algorithm's performance in real-world scenarios needs further testing.

Summary updated

BibTeX
@inproceedings{principi2013distributed,
  author = {Emanuele Principi and Stefano Squartini and Francesco Piazza and Danilo Fuselli and Maurizio Bonifazi},
  title = {{A distributed system for recognizing home automation commands and distress calls in the Italian language}},
  booktitle = {Interspeech},
  year = {2013},
}
Figure