Efficient three-stage pitch estimation for packet loss concealment

Xuejing Sun, Sameer GadreInterspeech 2010

TL;DR

This paper introduces a novel, three-stage pitch estimation algorithm designed for efficient packet loss concealment (PLC) in voice communication. By breaking down pitch estimation into progressively refined stages and optimizing for embedded systems like Bluetooth, it achieves significant computational savings (32% fewer cycles) while improving speech quality compared to existing methods.

What Problem Does This Paper Address?

In wireless and VoIP communications, lost data packets can lead to jarring clicks and pops, severely degrading speech quality and intelligibility. Packet Loss Concealment (PLC) aims to mitigate this by generating synthetic speech to fill the gaps. A crucial component of many PLC techniques, particularly waveform substitution methods, is accurate pitch estimation – determining the fundamental frequency of speech. However, existing pitch estimation algorithms are often too computationally intensive for resource-constrained devices like Bluetooth chips. This paper addresses the need for a low-complexity, yet accurate, pitch estimation algorithm suitable for real-time PLC on such embedded platforms.

Methods

The proposed algorithm employs a three-stage approach to pitch estimation:

  • Stage One: Coarse Pitch Estimation: This stage uses a narrow pitch range (e.g., 75-150 Hz) and a decimated input signal (factor of 2) to quickly identify a candidate pitch lag. The Normalized Cross-Correlation (NCC) metric is used, with an approximation for the denominator to reduce computational cost.
  • Stage Two: Pitch Multiples Refinement: This stage checks integer multiples of the pitch found in Stage One. This is crucial because the narrow range in Stage One might have identified a multiple of the true pitch. The algorithm efficiently searches for the best pitch candidate among these multiples.
  • Stage Three: Boundary Refinement: This stage focuses on minimizing discontinuities at the boundaries of the synthesized speech segment. It refines the pitch estimate by minimizing the Sum of Squared Differences (SSD) between the signal segments just before and after the lost packet, considering the candidate pitch lag. This stage is designed to improve the smoothness of the cross-fading process used in PLC.

The algorithm is optimized for embedded DSPs by leveraging efficient calculations and approximations.

Datasets

The paper uses clean speech files from the NOIZEUS dataset, which contains 30 IEEE sentences spoken by six speakers (three male, three female). These files were artificially corrupted by dropping frames of audio samples at random to simulate packet loss at rates of 5%, 10%, 15%, and 20%. Packet sizes of 30 and 60 samples at an 8 kHz sampling rate were used, corresponding to common Bluetooth packet types (HV3/EV3 and 2-EV3).

Key Insights and Ideas

  • Narrow Pitch Range Efficiency: Restricting the initial pitch search to a narrow range, especially when combined with signal decimation, dramatically reduces computational load without sacrificing essential information for PLC.
  • Multi-stage Refinement is Key: The three-stage approach effectively balances computational efficiency with accuracy. Stage Two handles potential multiples, while Stage Three directly addresses the critical issue of boundary discontinuities in synthesized speech.
  • Optimizing for Embedded Systems: The algorithm's design, including approximations and efficient calculations, is tailored for DSP processors commonly found in devices like Bluetooth chips, making it practical for real-world deployment.
  • Pitch Multiples Suffice for PLC: For PLC, the exact pitch isn't always necessary; pitch multiples often provide sufficient information for generating a plausible synthetic signal, simplifying the estimation process.

Key Results

  • Computational Savings: The proposed algorithm requires approximately 32% fewer instruction cycles compared to a similarly decimated G.711 Appendix I approach when implemented on a CSR Kalimba DSP.
  • Improved Speech Quality: Objective evaluation using the PESQ (Perceptual Evaluation of Speech Quality) MOS-LQO score showed that the proposed algorithm consistently outperforms the baseline G.711 system across various packet loss rates and packet sizes. The improvement was more significant with each additional stage of refinement.
  • Stage 1+3 Performance: Notably, for a 30-sample packet size, the system with only Stage One and Stage Three outperformed the baseline in all tested loss conditions.
  • Smaller Packet Size Advantage: All tested systems, including the proposed one, generally performed better with the smaller 30-sample packet size.

Limitations and Future Work

  • Limitations: The paper acknowledges that while the algorithm significantly reduces discontinuities, they can still remain a problem, particularly at higher loss rates. The current implementation uses a fixed overlap-add (OLA) length for efficiency, which might not be optimal for all scenarios.
  • Future Work: The authors suggest future work will involve more comprehensive evaluation of the algorithm in a real Bluetooth communication environment, specifically addressing scenarios prone to bursty errors. They also aim to explore further optimizations and potentially adaptive OLA lengths.

Summary updated

BibTeX
@inproceedings{sun2010efficient,
  author = {Xuejing Sun and Sameer Gadre},
  title = {{Efficient three-stage pitch estimation for packet loss concealment}},
  booktitle = {Interspeech},
  year = {2010},
}
Figure