TL;DR
This paper introduces a novel bio-radar system operating at 94 GHz for speech acquisition, overcoming the limitations of traditional microphones. To address the inherent noise in radar-captured speech, they propose a Minimum Mean-Square Error (MMSE) short-time spectral amplitude (STSA) estimation algorithm, demonstrating significant improvements in speech quality.
What Problem Does This Paper Address?
Traditional microphones, while ubiquitous, have significant drawbacks for speech acquisition: limited range (around 2 meters), poor directivity, weak noise compensation, and low sensitivity. This necessitates alternative methods. The authors' lab has developed a bio-radar system that leverages millimeter waves to detect speech by analyzing vibrations on the throat or chest. This bio-radar approach offers advantages like low attenuation, high directivity, wide bandwidth, and high sensitivity, potentially extending speech acquisition capabilities. However, the raw radar-captured speech is contaminated by various noises, including electromagnetic interference, circuit noise, and ambient noise, which degrade its intelligibility and audibility. This paper aims to tackle this challenge by enhancing the quality of radar-acquired speech.
Methods
The core of the paper lies in the proposed Log-spectra MMSE Radar Speech Enhancement Algorithm. This algorithm is designed to enhance radar-captured speech by minimizing the mean-square error of the log-spectral amplitude. The rationale is that human hearing is more sensitive to spectral amplitude than to the spectral component itself.
The process involves:
- Radar Speech Acquisition: A 94 GHz millimeter-wave radar system is used to capture speech signals by detecting throat vibrations.
- Noise Modeling: The received noisy speech signal
y(t)is modeled as the sum of clean speechx(t)and noisen(t). - MMSE Estimation: The algorithm aims to estimate the clean speech spectral amplitude
A_kfrom the noisy signal. It utilizes a distortion measure based on the mean-square error of the log-spectra:E{(log A_k - log A_k)^2}. - Optimal Spectral Amplitude Estimator: The algorithm employs an optimal spectral amplitude estimator derived from the MMSE principle, which is particularly beneficial for auditory perception.
- Algorithm Implementation: The derivation involves statistical modeling (Gaussian model) and integration to arrive at the final amplitude estimator.
Datasets
Not discussed in the paper. The experiments involved human volunteers speaking Mandarin sentences in a controlled environment.
Key Insights and Ideas
- Bio-radar as a Viable Speech Acquisition Method: The research demonstrates that millimeter-wave bio-radar can successfully detect speech signals, offering a promising alternative to traditional microphones.
- Log-spectra MMSE for Enhanced Audibility: Focusing on the log-spectral amplitude for enhancement is crucial because it aligns better with human auditory perception, leading to more natural-sounding enhanced speech.
- Effective Noise Reduction: The proposed MMSE STSA algorithm significantly reduces colored residual noise, outperforming traditional spectral subtraction and wavelet transform denoising methods.
- Potential for Wide Application: The successful enhancement of radar-acquired speech opens doors for integrating this technology into various communication devices, improving overall speech quality.
Key Results
- Subjective Evaluation (MOS): The enhanced radar speech from the proposed algorithm was perceived as more pleasant, with reduced residual noise and minimal speech distortion.
- Objective Evaluation (PESQ): Table 1 shows that the proposed algorithm consistently achieved higher PESQ scores compared to the original radar speech and other enhancement methods, especially at low Signal-to-Noise Ratios (SNRs). For instance, at -10 dB SNR, the proposed algorithm achieved a PESQ score of 1.26, significantly higher than the original speech's 0.38. At 10 dB SNR, it reached 3.75, outperforming others.
- Spectrogram Analysis: Figure 3 visually confirms the effectiveness of the proposed algorithm in reducing noise, particularly in the high-frequency regions, while preserving speech components.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the detected radar speech quality is lower than traditional microphone speech, with degraded intelligibility and audibility, necessitating the enhancement.
- The study mentions that radar wave illumination can be potentially hazardous, but emphasizes that the power levels used in their experiment were very low (100 mW, ~1% of a mobile phone's power) and that millimeter waves penetrate human tissue minimally.
- Future Work:
- The authors suggest that in future work, a larger size of radar antennas will be used to further improve the performance of the bio-radar system.
- The paper highlights the necessity of speech enhancement algorithms for the further improvement of radar speech quality and suggests their potential integration into telephones, microphones, or other electronic equipment.
Summary updated