Non-linear predictive vector quantization of feature vectors for distributed speech recognition

José Enrique García Laínez, Alfonso Ortega, Antonio Miguel, Eduardo LleidaInterspeech 2010 · DOI

Introduction

Distributed Speech Recognition (DSR) architectures decouple feature extraction from the recognition engine, allowing mobile clients to offload computationally intensive tasks to a back-end server. While this approach facilitates voice-enabled applications on resource-constrained devices, the transmission of acoustic features introduces a significant bandwidth bottleneck. Standard compression techniques, such as the ETSI 4.4 kbps standard, provide a baseline, but high-concurrency environments and usage-based data plans demand even lower bit-rates.

Previous research into Differential Vector Quantization (DVQ) utilized linear prediction to reduce redundancy. However, since speech production is inherently non-linear, linear models often fail to capture complex temporal correlations, leading to performance degradation at bit-rates below 2.1 kbps. This paper proposes a Predictive Vector Quantization scheme using a Multi-Layer Perceptron (PVQ-MLP) to exploit these non-linearities and achieve substantial bit-rate reductions with minimal impact on Word Error Rate (WER).

Methodology

The core of the proposed system is the PVQ-MLP architecture. Unlike traditional linear predictors that use a weighted sum of past frames, the PVQ-MLP employs an artificial neural network to predict the current feature vector.

MLP Architecture and Multichannel Input

The predictor is implemented as a feed-forward MLP with a single hidden layer containing seven neurons and a hyperbolic tangent sigmoid (tansig) transfer function. A critical innovation is the "multichannel" input strategy: the network predicts the current Mel-Frequency Cepstral Coefficient (MFCC), $C_i(t)$, by observing not only past values of that specific coefficient but also the past log-energy ($E$) of the signal.

Architecture of the MLP-based non-linear predictor incorporating log-energy inputs.

As shown in the architecture above, the model typically utilizes a history of three frames ($t-1$ to $t-3$). The inclusion of energy provides essential context regarding the signal's dynamics, which significantly improves prediction accuracy for lower-order MFCCs.

Quantization and Bit Allocation

Following prediction, the system calculates the residual error between the predicted and actual feature vectors. This residual is then quantized using Vector Quantization (VQ). To optimize performance at extremely low bit-rates, the authors refined the bit allocation across different MFCC coefficient groups, prioritizing those most critical for recognition accuracy.

Experimental Setup

The system was evaluated using the Aurora 4 corpus, a robust speech recognition benchmark based on the Wall Street Journal (WSJ0) task. The dataset includes 5,000-word vocabulary tasks across various training conditions (clean, multi-noise, multi-condition) and 14 test sets featuring diverse additive noises and microphone distortions. All tests utilized the 8 kHz database version and the ETSI Advanced Front-End (AFE).

Results and Discussion

The PVQ-MLP system demonstrates a significant advantage over both standard VQ and linear predictive methods (DVQ).

Bit-rate vs. Accuracy Trade-off

The primary achievement of this work is the reduction of the bit-rate to 1.8 kbps—a 66% decrease compared to the ETSI 4.4 kbps standard—while maintaining a WER degradation of less than 5% relative to an unquantized system.

Mean Word Error Rate (WER) across various bit-rates for different quantization schemes.

As illustrated in the performance curves, standard VQ begins to fail rapidly below 3.5 kbps. In contrast, the PVQ-MLP maintains stability down to approximately 1.4 kbps. At 1.8 kbps, the proposed non-linear method matches the recognition performance of standard VQ at 3.5 kbps and the ETSI standard at 4.4 kbps.

Insights on Temporal Correlation

The study found that looking back more than two or three frames provided diminishing returns, suggesting that the most relevant correlations in speech features are short-lived. Furthermore, the non-linear approach was particularly effective for lower-order coefficients ($C_0, C_1$), which are the most vital for robust recognition.

Conclusion

By transitioning from linear to non-linear prediction via an MLP, the authors successfully pushed the boundaries of low-bit-rate DSR. The PVQ-MLP scheme achieves a 1.8 kbps bit-rate with negligible loss in accuracy, offering a viable solution for high-density voice applications. Future work will address the system's robustness against channel errors and the computational overhead on mobile hardware.

Summary updated

BibTeX
@inproceedings{lainez2010non,
  author = {José Enrique García Laínez and Alfonso Ortega and Antonio Miguel and Eduardo Lleida},
  title = {{Non-linear predictive vector quantization of feature vectors for distributed speech recognition}},
  booktitle = {Interspeech},
  year = {2010},
  doi = {10.21437/interspeech.2010-650},
}
Figure