A MFCC-based CELP speech coder for server-based speech recognition in network environments

Gil Ho Lee, Jae Sam Yoon, Hong Kook KimInterspeech 2005

TL;DR

This paper introduces a novel CELP speech coder that uses Mel-Frequency Cepstral Coefficients (MFCCs) for spectral envelope representation, aiming to improve speech recognition accuracy in network environments. By employing predictive vector quantization (PVQ) with a safety-net scheme, they achieve a comparable speech quality to G.729 at a slightly lower bit-rate, while expecting significant gains in Automatic Speech Recognition (ASR) performance.

What Problem Does This Paper Address?

Traditional speech coders, like those based on Linear Predictive Coding (LPC), optimize for speech quality. However, the reconstructed speech often suffers from degraded performance when used as input for Automatic Speech Recognition (ASR) systems. This is because the spectral envelope parameters optimized for human perception (like LPC) are not necessarily optimal for ASR features. Mel-Frequency Cepstral Coefficients (MFCCs) are known to be more effective for ASR. The challenge is to develop a speech coder that can efficiently represent speech using MFCCs at a low bit-rate, suitable for network environments, without sacrificing too much speech quality, and crucially, to improve ASR performance.

Methods

The proposed system is a Code-Excited Linear Prediction (CELP) coder, similar in structure to the ITU-T G.729 standard. The key innovation lies in using MFCCs for spectral envelope representation instead of LPC.

  • MFCC Extraction: Standard MFCC extraction is performed, involving pre-processing, windowing, FFT, mel-filtering, log-scaling, and Discrete Cosine Transform (DCT).
  • MFCC to LPC Conversion: For the CELP synthesis filter, a conversion process from MFCCs back to LPC coefficients is implemented. This allows the CELP decoder to reconstruct speech while the MFCCs themselves are transmitted for ASR.
  • Predictive Vector Quantization (PVQ): To reduce the bit-rate, the authors exploit the interframe correlation of MFCCs. PVQ is used to predict current MFCCs based on past quantized MFCCs.
  • Safety-Net PVQ: To mitigate the impact of channel errors and frame erasures, a safety-net scheme is introduced. This combines PVQ with a memoryless Vector Quantizer (VQ). A decision is made based on Euclidean distance to select either PVQ or memoryless VQ for each subvector, aiming to reduce error propagation.
  • Multi-stage VQ: For the 12-dimensional MFCC subvector (C2), a multi-stage VQ is employed for efficient search.

Datasets

The paper mentions using a speech database for training and evaluation.

  • Training: 3,200 frames collected from utterances spoken by 2 males and 2 females were used to measure interframe correlations.
  • VQ Training: A larger dataset of 172,800 frames from American, English, and Korean speakers was used for training the proposed VQ.
  • VQ Evaluation: 48,400 frames from the same speakers were used for evaluating the VQ.
  • Performance Evaluation: 64 sentences spoken by four male and four female speakers were used for perceptual evaluation.

Key Insights and Ideas

  • MFCCs for ASR-Friendly Coding: The core idea that MFCCs, known for ASR, can be directly used in a speech coder for improved recognition performance is a significant departure from traditional LPC-based approaches.
  • Interframe Correlation for Efficiency: Exploiting the high interframe correlation of MFCCs is crucial for achieving low bit-rates through predictive coding. The paper quantifies this correlation, showing that MFCCs are highly predictable from previous frames.
  • Safety-Net for Robustness: The safety-net PVQ approach is a clever way to balance the efficiency of predictive coding with the robustness needed for real-world network conditions, where channel errors are common.
  • Trade-off between Quality and ASR: The paper highlights that while optimizing for speech quality is important, optimizing for ASR performance might require different feature representations and quantization strategies, even if it means a slight trade-off in perceptual quality.

Key Results

  • Bit-Rate: The proposed MFCC-based CELP coder operates at 8.7 kbps, which is comparable to the 8 kbps G.729 standard.
  • MFCC Quantization: The MFCCs are quantized using 25 bits per frame, a reduction of 19 bits compared to the previous SVQ approach (which required 44 bits).
  • Speech Quality (PESQ): Under noiseless conditions, the proposed coder achieved a comparable or slightly better MOS score than G.729. Specifically, with optimal prediction coefficients (a2 between 0.85 and 0.95), the proposed coder was about 0.02 MOS higher than G.729.
  • Noisy Channel Performance: The coder demonstrated robustness to channel errors, with performance degradation that was manageable and comparable to G.729's frame erasure concealment.
  • Expected ASR Improvement: While not directly measured in this paper, the authors strongly expect that ASR systems using the transmitted MFCCs will achieve higher recognition accuracy than those using G.729's reconstructed speech.

Limitations and Future Work

  • MFCC to LPC Conversion Complexity: The conversion from MFCC to LPC introduces an additional processing step in the decoder, which might add computational complexity.
  • Optimal Parameter Selection: The paper focuses on finding optimal prediction coefficients (a1, a2) for PVQ. Further research could explore more adaptive or data-driven methods for selecting these parameters.
  • Direct MFCC Transmission for ASR: The primary benefit is expected for ASR. Future work would involve integrating this coder into a full server-based ASR system and empirically demonstrating the claimed ASR performance gains.
  • Further Bit-Rate Reduction: While 8.7 kbps is good, further exploration into more aggressive quantization techniques for MFCCs could potentially lead to even lower bit-rates.

Summary updated

BibTeX
@inproceedings{lee2005mfcc,
  author = {Gil Ho Lee and Jae Sam Yoon and Hong Kook Kim},
  title = {{A MFCC-based CELP speech coder for server-based speech recognition in network environments}},
  booktitle = {Interspeech},
  year = {2005},
}
Figure