Analysis of Temporal Resolution in Frequency Domain Linear Prediction

Sriram Ganapathy, Hynek HermanskyInterspeech 2012 · DOI

Abstract

Frequency Domain Linear Prediction (FDLP) provides a robust framework for modeling the temporal envelopes of speech, offering potential advantages for Automatic Speech Recognition (ASR) in noisy environments. However, standard FDLP techniques often suffer from inconsistent temporal resolution across analysis windows. This paper presents a "High Resolution" (FDLP-HR) approach that optimizes window functions, prediction methods, and signal padding to achieve superior temporal precision. Evaluation on the TIMIT and CTS datasets demonstrates that FDLP-HR features achieve a 10% relative improvement in phoneme recognition under noisy conditions compared to standard MFCC features.

Introduction

ASR systems frequently experience performance degradation when speech is corrupted by background noise. Standard features like Mel-Frequency Cepstral Coefficients (MFCCs) are sensitive to noise because it significantly alters the low-energy regions of the signal. FDLP offers a more robust alternative by modeling the Hilbert envelope—the smooth outline of a signal's amplitude over time. Because FDLP focuses on high-energy peaks, it is less susceptible to noise masking. The primary challenge addressed in this work is the "blurriness" or inconsistent resolution of standard FDLP, which typically captures detail well only in the center of an analysis window.

Enhancing Temporal Resolution

The core of this research involves analyzing the "critical time-span," defined as the minimum distance required between two temporal peaks for the model to distinguish them as separate entities.

Illustration of temporal resolution in FDLP showing resolved (a, b) and unresolved (c, d) impulses.

To improve this resolution and ensure consistency across the analysis window, the authors proposed several modifications to the FDLP pipeline:

  • Linear Prediction on DCT: Instead of performing linear prediction in the time domain, it is performed on the Discrete Cosine Transform (DCT) of the signal to obtain a temporal envelope.
  • Gaussian Windows: Applying a Gaussian window to the DCT sequence provides superior temporal resolution compared to standard rectangular or Hamming windows.
  • Least-Squares Linear Prediction: Swapping the standard auto-correlation method for least-squares LP better preserves sharp temporal peaks.
  • Symmetric Padding: To resolve the drop in resolution at window boundaries, the signal is symmetrically padded with approximately 32ms of mirrored data, which is discarded after processing.
  • Optimized Model Order: Increasing the number of poles allows for finer detail, though the authors identified a "sweet spot" beyond which the model begins to fit the noise itself.

Experimental Results

The FDLP-HR features were rigorously tested against standard MFCC baselines using the TIMIT and Conversational Telephone Speech (CTS) datasets.

Performance in Noise

On the TIMIT dataset, which included various noise types (Babble, Restaurant, Subway) at SNRs from 0dB to 20dB, FDLP-HR achieved an average 10% relative improvement in Phoneme Error Rate (PER). The high-resolution approach was particularly beneficial for transient sounds like plosives ('p', 't', 'k') and nasals ('m', 'n'), which rely on fine-grained temporal transitions.

Telephone Speech Robustness

In CTS experiments representing real-world telephone channel conditions, the high-resolution features provided a 5% relative improvement. This indicates that the technique generalizes well to channel distortion in addition to additive noise.

Conclusion

By addressing the inherent resolution limitations of FDLP, the proposed FDLP-HR framework significantly enhances ASR robustness. The ability to maintain sharp temporal peaks in noisy conditions ensures a better match between clean training data and corrupted test signals. Future research will focus on integrating these features with other robust front-end techniques like spectral subtraction to further improve performance in extreme noise environments.

Summary updated

BibTeX
@inproceedings{ganapathy2012analysis,
  author = {Sriram Ganapathy and Hynek Hermansky},
  title = {{Analysis of Temporal Resolution in Frequency Domain Linear Prediction}},
  booktitle = {Interspeech},
  year = {2012},
  doi = {10.21437/interspeech.2012-497},
}
Figure