Multi-Microphone Adaptive Noise Cancellation for Robust Hotword Detection

Yiteng Huang, Turaj Z. Shabestary, Alexander Gruenstein, Li WanInterspeech 2019

TL;DR

This paper extends a novel adaptive noise cancellation (ANC) technique, the "hotword cleaner," to multi-microphone setups for more robust hotword detection. By leveraging the unique properties of hotwords (short, leading phrases), the system effectively cancels noise without requiring complex speech enhancement or mask estimation, leading to significantly reduced false rejection rates.

What Problem Does This Paper Address?

Hotword detection, the technology that allows devices like smart speakers and smartphones to wake up upon hearing specific phrases (e.g., "Hey Google"), is crucial for voice assistant functionality. However, this always-on system faces significant challenges in noisy environments, where background sounds can easily trigger false alarms or prevent the hotword from being recognized. Traditional noise cancellation methods often struggle with complex, speech-like noise (like TV audio) and can introduce unwanted speech distortion. Existing deep learning approaches have improved robustness, but there's a continuous need for more efficient and effective noise cancellation, especially for resource-constrained devices. This paper aims to improve the robustness of hotword detection in noisy conditions by generalizing a previously proposed dual-microphone ANC method to handle more microphones.

Methods

The core of this work is the generalization of the "hotword cleaner" algorithm to multi-microphone systems, termed "McCleaner." Unlike traditional ANC that aims for distortionless speech, the hotword cleaner prioritizes noise cancellation by exploiting two key hotword characteristics: they are the first words in a voice query and are short. This allows the system to avoid complex speech-noise mask estimation.

The McCleaner operates in the frequency domain using the Short-Time Fourier Transform (STFT). It models the microphone signals as a combination of the desired speech and additive noise. The key innovation lies in using a deferred filter coefficient approach. Instead of using the most up-to-date filter coefficients, it uses coefficients that were computed with a time lag. This allows the system to adapt to changing noise conditions while still effectively canceling noise that was present when the filter was computed.

The algorithm's adaptation relies on estimating the covariance of the noise signals across microphones. This estimation is done recursively, and importantly, it continues even during speech presence. The filter coefficients are then computed by inverting this noise covariance matrix. For multi-microphone setups (M > 2), the authors use a Singular Value Decomposition (SVD) based pseudoinverse method to handle potential singularity of the covariance matrix.

Datasets

The authors collected re-recorded audio data for evaluation. This dataset consists of:

  • Far-Field Clean: Utterances in a clean acoustic environment.
  • TV Noise: Utterances with background TV noise, recorded using a 3-microphone array.
  • Negative: A dataset of TV audio recorded from a far-field Google Home device in previous studies, used for evaluating false alarms.

The data was sampled at 16 kHz with 16 bits.

Key Insights and Ideas

  • Hotword Properties as a Feature: The paper strongly emphasizes that the inherent properties of hotwords (short duration, leading phrase) are not just incidental but can be actively exploited to design more efficient noise cancellation systems. This is a departure from general-purpose speech enhancement.
  • Deferred Filter Coefficients for Dynamic Noise: The use of deferred filter coefficients is a clever way to balance adaptation to changing noise environments with the need for effective cancellation. It allows the system to "look back" at a slightly older noise profile for cancellation, which is sufficient for tracing ambient noise.
  • Multi-Microphone Advantage Beyond Beamforming: While beamforming is a common multi-microphone technique, this work shows that even for noise cancellation (not necessarily distortionless speech), adding more microphones can significantly improve performance by providing more reference signals for noise prediction and cancellation.
  • Simplicity and Efficiency: The McCleaner algorithm, despite its effectiveness, is presented as having a simple logic and being computationally efficient, making it suitable for embedded devices.

Key Results

The paper demonstrates that the McCleaner algorithm significantly outperforms a baseline system, especially in noisy conditions.

  • On the TV-noise dataset, the false rejection (FR) rate was reduced from approximately 75% to below 17% at a comparable operating point.
  • By increasing the number of microphones from 2 to 3, the FR rate was further reduced by a relative 33.5%. This improvement was observed in both cleaner-only and hybrid system configurations.
  • Qualitative analysis showed that the 3-mic McCleaner produced a perceptibly lower and smoother residual noise compared to the 2-mic version, aiding in the recognition of challenging phonemes.

Limitations and Future Work

  • Covariance Matrix Singularity: While the SVD-based pseudoinverse method is used to address potential singularity of the noise covariance matrix with M > 2, the paper notes that for M=2 and moderate L, the matrix is presumed non-singular, and a Fast-RLS method was used previously. The robustness of the SVD method for very large M or complex noise scenarios might warrant further investigation.
  • Voice Activity Detector (VAD) Robustness: The paper acknowledges that VAD can be problematic when additive noise contains speech. While their scheme adapts continuously, the reliance on accurate noise estimation is implicit.
  • Future Work: The paper suggests that the cleaner-only strategy is computationally efficient and reduces false alarms, while the hybrid strategy minimizes latency and handles quiet corner cases. Further exploration of these integration strategies and their trade-offs in various real-world scenarios would be valuable. The authors also implicitly suggest that further improvements might come from exploring different microphone array geometries or more advanced adaptation schemes.

Summary updated

BibTeX
@inproceedings{huang2019multi,
  author = {Yiteng Huang and Turaj Z. Shabestary and Alexander Gruenstein and Li Wan},
  title = {{Multi-Microphone Adaptive Noise Cancellation for Robust Hotword Detection}},
  booktitle = {Interspeech},
  year = {2019},
}
Figure