Quality Evaluation of Reverberation in Audioband Speech Signals

Amaro A. de Lima, Sergio L. Netto, Luiz W. P. Biscainho, Fabio P. Freeland, Bruno C. Bispo, Rafael A. de Jesus, Ronald W. Schafer, Amir Said, Bowon Lee, Ton KalkerICETE (Selected Papers) 2008 · DOI

TL;DR

This paper tackles the challenge of evaluating reverberation in high-quality audioband speech signals for telepresence systems. The authors propose and validate a new objective metric, ALcons, which shows a strong correlation with subjective listener scores, offering a more effective way to assess speech quality degradation due to room acoustics.

What Problem Does This Paper Address?

Modern telepresence systems aim to provide a highly immersive meeting experience, which includes delivering high-quality audio. However, the acoustic properties of the rooms where participants are located can significantly degrade speech quality through reverberation. This phenomenon, characterized by echoes and a prolonged decay of sound, reduces speech intelligibility and can negatively impact user satisfaction. While existing methods for speech quality assessment exist, they are often designed for narrower bandwidths or don't specifically address the nuances of reverberation in high-fidelity audio. This paper addresses the need for a robust and accurate method to objectively measure the impact of reverberation on audioband speech quality, bridging the gap between subjective listener perception and objective technical metrics.

Methods

The paper explores several approaches to modeling and evaluating reverberation:

  • Reverberation Modeling:
    • Early Reflections: Modeled using the "image method," which simulates reflections by creating virtual sound sources based on room geometry.
  • Late Reverberation: * Schroeder's Reverberator: Utilizes comb and all-pass comb filters to create a diffuse sound field.
  • Feedback Delay Networks (FDN): A more generalized approach using multiple delay lines and feedback paths to generate complex reverberation.
  • Gardner's Late Reverberator: Employs nested all-pass filters and considers different schemes for small, medium, and large rooms.
  • Modified Gardner's Full Reverberator: A proposed modification to Gardner's approach to achieve a smoother transition between early reflections and late reverberation, improving naturalness.
  • Subjective Evaluation: A listening test was conducted where participants rated the quality of speech signals degraded by various levels of artificial reverberation. Mean Opinion Scores (MOS) were collected.
  • Objective Evaluation:
    • Standard Metrics: Existing objective quality assessment tools like PESQ, W-PESQ, P.563, PEAQ (Basic and Advanced), and Rnonlin were evaluated for their correlation with subjective MOS.
  • ALcons (Articulation Loss of Consonants): A novel objective metric proposed by Peutz, which is adapted and validated in this work. ALcons estimates the percentage of consonants that are unintelligible due to reverberation and distance. The paper details methods (Algorithm V1 and V2) to estimate the necessary parameters (room volume, source-microphone distance, and reverberation time) from the Room Impulse Response (RIR).

Datasets

The study used artificial Room Impulse Responses (RIRs) generated to simulate specific reverberation times (T60) ranging from 200 ms to 700 ms. These RIRs were convolved with original, high-quality speech recordings (Brazilian-Portuguese sentences) from four different speakers. The original recordings were made in a small, acoustically treated room at various source-microphone distances. No new benchmark datasets were introduced; rather, existing speech signals were processed to create the test material.

Key Insights and Ideas

  • Reverberation is a critical impairment: The study confirms that reverberation significantly degrades the perceived quality of audioband speech, with longer reverberation times leading to lower subjective scores.
  • Existing metrics fall short: Standard objective quality assessment tools, primarily designed for narrower bandwidths or different impairments, show only moderate correlations with subjective reverberation quality.
  • ALcons shows promise: The ALcons metric, when adapted and applied to audioband speech, demonstrates a remarkably high correlation with subjective MOS, suggesting it's a powerful tool for assessing reverberation impact.
  • T60 is important, but not the only factor: While reverberation time (T60) is a primary determinant of quality degradation, the study acknowledges that other factors also influence subjective perception.
  • Modeling reverberation is complex: The paper reviews various models for early and late reverberation, highlighting the trade-offs and complexities in accurately simulating room acoustics.

Key Results

  • Subjective listening tests showed a clear decrease in Mean Opinion Scores (MOS) as reverberation time (T60) increased. For instance, signals with T60 of 196 ms received MOS scores similar to "anchor" (original, non-anechoic) signals, while longer T60 values resulted in significantly lower scores.
  • Standard objective quality assessment methods showed correlations with MOS ranging from 0.23 (PEAQ Adv.) to 0.84 (PESQ).
  • The proposed ALcons metric, using estimated room parameters from RIRs, achieved a high correlation of p = 92% with subjective MOS, significantly outperforming other objective methods.
  • The paper validates two algorithms (V1 and V2) for estimating room volume from RIRs, showing a strong linear relationship with theoretical values, which is crucial for ALcons calculation.

Limitations and Future Work

  • Limitations:
    • The study acknowledges that T60 is not the sole determinant of subjective quality. Other aspects of the RIR and signal processing might also play a role.
  • Some of the objective metrics evaluated are designed for speech-band (4 or 8 kHz) signals, making direct comparisons with audioband subjective tests require careful consideration.
  • Informal listening tests indicated that the FDN method might lose naturalness above certain reverberation times, leading to the use of Gardner's modified method for longer T60 values.
  • Future Work:
    • The authors suggest that the ALcons measure can be efficiently employed for reverberation assessment in audioband speech signals, implying its potential for real-time quality monitoring in telepresence systems.
  • Further research could explore the influence of other RIR characteristics beyond T60 on subjective quality.
  • Investigating the performance of ALcons in more diverse acoustic environments and with different types of audio content could be beneficial.

Summary updated

BibTeX
@inproceedings{lima2008quality,
  author = {Amaro A. de Lima and Sergio L. Netto and Luiz W. P. Biscainho and Fabio P. Freeland and Bruno C. Bispo and Rafael A. de Jesus and Ronald W. Schafer and Amir Said and Bowon Lee and Ton Kalker},
  title = {{Quality Evaluation of Reverberation in Audioband Speech Signals}},
  booktitle = {ICETE (Selected Papers)},
  year = {2008},
  doi = {10.1007/978-3-642-05197-5\_28},
}
Figure