A media-specific FEC based on huffman coding for distributed speech recognition

Young Han Lee, Hong Kook KimInterspeech 2009

TL;DR

This paper introduces a novel media-specific Forward Error Correction (FEC) method for Distributed Speech Recognition (DSR) that significantly reduces word error rates in noisy environments. By intelligently protecting the most crucial parts of speech features (MFCCs) and using a hybrid Huffman coding approach for compression, the system achieves substantial performance gains without increasing bitrates.

What Problem Does This Paper Address?

The increasing prevalence of mobile and portable devices has fueled the demand for speech recognition. However, the high computational cost of speech recognition makes it challenging for these devices. Distributed Speech Recognition (DSR) offers a solution by offloading the heavy lifting to a server, with the device performing front-end feature extraction and compression. A critical challenge in DSR, especially over wireless networks, is the degradation of speech recognition performance due to channel errors that corrupt the transmitted features (like MFCCs). While Forward Error Correction (FEC) can mitigate this, simply adding more error protection bits increases the bitrate, which is undesirable for bandwidth-constrained systems. This paper aims to address this by developing an efficient FEC method that targets the most important speech features and uses advanced compression techniques to maintain a low bitrate.

Methods

The core of this work lies in two main innovations:

  • Media-Specific FEC: The authors first analyze the importance of different subvectors within the Mel-Frequency Cepstral Coefficients (MFCCs) for DSR performance. Through experiments on the Aurora 4 database, they identify the first subvector (specifically, C1 and C2) as being the most critical. Their media-specific FEC method then focuses on protecting this particular subvector. To recover a lost frame, the system uses information from the previous frame, introducing a small algorithmic delay.
  • Hybrid Huffman Coding: To avoid increasing the bitrate with the added FEC information, the paper proposes a hybrid Huffman coding approach. This method combines two strategies:
  • Class-Dependent Huffman Coding: Recognizes that the entropy (and thus compressibility) of MFCCs varies with the voicing class of the speech (e.g., voiced vs. unvoiced). Different Huffman tables are created for different voicing classes.
  • Subvector-Wise Huffman Coding: Further exploits the fact that different subvectors within the MFCCs have varying levels of redundancy. Separate Huffman tables are designed for each subvector. The "hybrid" aspect involves applying differential coding to both class-dependent and subvector-wise Huffman coding, leading to the smallest entropy and thus the most efficient compression.

Datasets

  • Aurora 4 Large Vocabulary Database: This is the primary dataset used for both training and evaluation. It's a standard benchmark for DSR systems and is used here to evaluate performance under various noisy conditions. The authors used a subset of the database comprising 166 utterances under seven different noise conditions (clean, car, babble, street traffic, airport, restaurant, and train station).

Key Insights and Ideas

  • The First MFCC Subvector is King: The most surprising and impactful insight is the disproportionate importance of the first MFCC subvector (C1, C2) for DSR performance. Protecting this single component significantly mitigates the impact of frame loss.
  • Entropy Varies with Context: The paper highlights that speech features are not uniformly redundant. Their compressibility (entropy) depends on both the voicing class of the speech and the specific subvector being considered. This observation is key to developing more efficient compression.
  • Hybrid Compression is Better: Combining class-dependent and subvector-wise Huffman coding, and then applying differential coding to these, yields the most significant bit-rate reduction. This suggests that exploiting multiple levels of redundancy is crucial for optimal compression.
  • FEC and Compression Can Coexist: The paper demonstrates that it's possible to implement effective error protection (FEC) without sacrificing compression efficiency, a critical balance for DSR systems.

Key Results

The proposed media-specific FEC method, combined with the hybrid Huffman coding, achieved significant improvements:

  • Word Error Rate (WER) Reduction: The system demonstrated relative average WER reductions of 9.03% to 17.81% compared to a standard DSR system without FEC, across various frame loss rates (5%, 10%, 20%, and 50%).
  • Bitrate Efficiency: The hybrid Huffman coding method achieved an average bit reduction of approximately 6.35 bits/frame, ensuring that the added FEC information did not lead to an overall bitrate increase.

Limitations and Future Work

  • Algorithmic Delay: The proposed FEC method introduces an algorithmic delay of about one frame length to recover lost information. While this is deemed acceptable due to the significant WER reduction, it's a trade-off to consider.
  • Focus on Single Frame Loss: The FEC mechanism is primarily designed to recover from single frame losses. The paper doesn't extensively detail performance under more severe or multi-frame loss scenarios, although the results under higher frame loss rates (20%, 50%) show continued benefits.
  • Future Work: The authors suggest that further exploration of hybrid Huffman coding combinations and potentially more sophisticated FEC schemes could lead to even greater performance gains. The paper also implicitly points towards the need for adaptive FEC strategies that can adjust protection levels based on real-time channel conditions.

Summary updated

BibTeX
@inproceedings{lee2009media,
  author = {Young Han Lee and Hong Kook Kim},
  title = {{A media-specific FEC based on huffman coding for distributed speech recognition}},
  booktitle = {Interspeech},
  year = {2009},
}
Figure