TL;DR
This paper introduces a novel method for accurately separating desired speech (foreground speech) from unwanted background noise and other speech (background degradation) in mobile phone recordings. By analyzing a specially filtered version of the audio signal, the authors achieve robust segmentation, significantly improving the performance of downstream speech processing tasks like vowel detection.
What Problem Does This Paper Address?
In today's wireless communication landscape, speech processing systems often receive audio from diverse and noisy environments. This means the desired speech from a specific speaker (foreground speech) is frequently mixed with various forms of background noise, including other people talking, vehicle sounds, and more. This "background degradation" can be quite similar to foreground speech, making it difficult for traditional methods to distinguish between them. This challenge is particularly acute for applications like voice-activated systems or speech recognition, where accurate identification of the desired speech is crucial for reliable operation. Existing methods often struggle when the background degradation is non-stationary or speech-like, leading to errors.
Methods
The core of this paper's contribution lies in its use of Zero Frequency Filtering (ZFF). This technique, applied to the incoming audio signal, effectively isolates and emphasizes the signal components around zero frequency. The authors then analyze features derived from this Zero Frequency Filtered Signal (ZFFS):
- Normalized First Order Autocorrelation Coefficient: This feature captures the periodicity of the ZFFS. It's found to be significantly higher for foreground speech compared to background degradation.
- Strength of Excitation: This feature, related to the slope of the ZFFS at positive zero crossings, also shows a marked difference between foreground speech and background.
These two features are combined into a single "evidence" signal. This evidence signal is then processed using an endpoint detection algorithm with carefully defined lower and upper thresholds to identify the boundaries of the foreground speech.
Datasets
The speech data used in this work was collected for an isolated word recognizer (IWR) task. Users called a voice server from mobile phones in various acoustic environments. The authors specifically mention using 60 files that exhibited significant background noise, including both speech-like and non-speech-like degradation (e.g., bird chirping, car horns, coughing). These files were manually labeled to serve as a reference for evaluating the proposed segmentation method.
Key Insights and Ideas
- ZFF as a powerful preprocessor: The paper demonstrates that ZFF can effectively extract discriminatory information between foreground speech and complex background degradation, even when traditional features like energy and ZCR show little difference.
- Distinct ZFFS characteristics: The ZFFS of foreground speech exhibits a more pronounced periodic behavior and higher excitation strength compared to background speech or noise, providing a clear signal for segmentation.
- Combined features for robustness: By combining the normalized autocorrelation coefficient and strength of excitation, the authors create a robust feature that effectively separates foreground speech from various types of background.
- Improved downstream task performance: The proposed segmentation method significantly reduces spurious detections in tasks like Vowel-like Region Onset Point (VLROP) detection, especially in challenging, speech-like background conditions.
Key Results
The proposed foreground speech segmentation method achieved high accuracy, with the average deviation for begin and end points of segmented foreground speech being 15.8 ms and 20.64 ms, respectively, compared to manually labeled segments.
Crucially, when applied to VLROP detection, the method led to a dramatic reduction in spurious VLROPs. Table 2 shows that without foreground segmentation, the error rate was 527% (likely due to a high number of spurious detections). With the proposed segmentation, the error rate dropped to 29%, primarily due to missed VLROPs within the foreground speech itself, rather than spurious ones in the background.
Limitations and Future Work
The authors acknowledge that the performance of their method degrades when the background speaker is very close to the sensor, as their acoustic characteristics become too similar to the foreground speech. They suggest that future work should focus on this specific scenario. Additionally, they propose exploring other features derived from the ZFFS for further improvements in foreground speech segmentation.
Summary updated