MONICA2: Mobile Neural Voice Command Assistants towards Smaller and Smarter

Yoonseok Hong, Shounan An, Sunwoo Im, Jaegeon Jo, Insoo OhAAAI 2022 · DOI

Abstract

Modern mobile games, particularly Massively Multiplayer Online Role-Playing Games (MMORPGs), are resource-intensive, often leaving little room for auxiliary features like voice command interfaces. MONICA2 is a highly optimized, on-device automatic speech recognition (ASR) system designed to address this. By applying cross-layer parameter sharing to the Conformer architecture, the system reduces parameter counts by 58% and memory footprint to 12.8MB, enabling real-time voice commands with minimal accuracy loss.

Introduction

Massively Multiplayer Online Role-Playing Games (MMORPGs) on mobile devices are notorious "resource hogs," frequently consuming over 1GB of RAM and saturating GPUs. While voice command interfaces could significantly enhance user experience—allowing players to trigger actions while "hands-busy"—standard high-performance ASR models are typically too large and computationally expensive to run alongside heavy game engines. MONICA2 bridges this gap by creating an ASR system that is "smaller and smarter": compact enough to fit in limited mobile memory while remaining accurate for specialized game commands.

Methods

The MONICA2 architecture optimizes the Conformer model, which integrates the global context-awareness of Transformers with the local feature extraction of Convolutions.

The MONICA2 pipeline and internal architecture, highlighting the cross-layer parameter sharing and intent classification components.

Key innovations include:

  • Cross-Layer Parameter Sharing (CLPS): Inspired by the ALBERT model, all 16 encoder blocks share a single Multi-Head Self-Attention Module (MHSAM), drastically reducing unique parameters.
  • Simplified Decoder: A traditional RNN-based decoder is replaced with a Linear Projection layer and Connectionist Temporal Classification (CTC) loss to streamline inference.
  • N-way Naive Bayesian Intent Classifier (NBIC): To map recognized text to game actions (e.g., "start auto-battle"), a lightweight classifier uses frequency-based vectors from the top-5 tokens of each frame.
  • Mobile-First Pre-processing: The system utilizes the CKFFT library for efficient Mel-spectrogram extraction on mobile CPUs.

Key Results and Insights

Performance profiling on a Samsung Galaxy S9 demonstrated that MONICA2 achieves a "sweet spot" between efficiency and accuracy. Sharing weights across 16 layers resulted in only a marginal (0.5%–0.6%) increase in Word Error Rate (WER), suggesting that the structure of the attention mechanism is more critical than unique weights per layer for this task.

Performance comparison of MONICA2 against other state-of-the-art models on the Librispeech benchmark and mobile efficiency metrics.

As shown in the results table, MONICA2 achieved:

  • Model Size: 58% fewer parameters than standard Conformer-CTC.
  • Memory Footprint: Only 12.8MB, a significant reduction from the 27MB required by the previous MONICA version.
  • Inference Speed: A Real-Time Factor (RTF) of 0.013, processing a 3-second command in approximately 30ms.
  • Accuracy: A 6.1% WER on Librispeech test-clean and an 88% success rate in identifying game-specific intents using the NBIC approach.

Conclusion

MONICA2 demonstrates that for specialized tasks like mobile game commands, a slight trade-off in general transcription accuracy is acceptable for massive gains in memory efficiency. By focusing on intent classification rather than perfect raw transcription, the system remains robust even when audio input is imperfect. Future work will explore more sophisticated lightweight Natural Language Understanding (NLU) components and test the system's performance in noisier real-world environments.

Summary updated

BibTeX
@inproceedings{hong2022monica,
  author = {Yoonseok Hong and Shounan An and Sunwoo Im and Jaegeon Jo and Insoo Oh},
  title = {{MONICA2: Mobile Neural Voice Command Assistants towards Smaller and Smarter}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21719},
}
Figure