Towards Improving Speaker Distance Estimation through Generative Impulse Response Augmentation

1. Problem Statement

The paper addresses the Speaker Distance Estimation (SDE) challenge, which is critical for applications like smart speakers, teleconferencing, hearing aids, and spatial audio. The core problem is that SDE models often suffer from limited training data, particularly in diverse acoustic environments. The specific challenge at ICASSP 2025 (GenDARA) focuses on two tasks:

1. Data Augmentation: Generating synthetic Room Impulse Responses (RIRs) to supplement sparse datasets.
2. Model Improvement: Fine-tuning SDE models using these augmented RIRs to improve distance estimation accuracy.

The goal is to determine if generative RIR data can effectively bridge the gap between limited real-world data and the robustness required for accurate distance prediction across varying speaker-listener distances.

2. Methodology

The authors propose a two-stage pipeline involving RIR generation and SDE model fine-tuning.

Task 1: Generative RIR Augmentation

• Base Model: The team utilized the open-source FastRIR (Fast Diffuse Room Impulse Response Generator), a conditional Generative Adversarial Network (GAN).
• Modifications: To align with the challenge constraints, the FastRIR architecture was modified to:
  • Condition solely on speaker and listener locations, removing room geometry conditioning.
  • Generate 1-second RIRs at a 32 kHz sampling rate.
  • Adapt input features to encode only distance-related parameters.
  • Adopt the MESH2IR representation scheme to ensure consistent energy distribution across distances.
• Training Strategy:
  • Pre-training: The network was pre-trained on 100,000 RIRs from the diverse 3D environments of the GWA dataset.
  • Fine-tuning: Separate fine-tuning processes were conducted for the Treble and GWA enrollment datasets. This was necessary because the two datasets use distinct simulation methodologies.
  • Data Split: For each fine-tuning phase, 80% of the enrollment RIRs were used for training, and 20% for validation/checkpoint selection.

Task 2: SDE Model Optimization

• Data Generation: Approximately 1 million RIRs were generated by sampling various speaker and listener positions.
• Quality Control (Filtering): Due to noise in the tails of generated RIRs leading to implausible reverberation times, a rigorous quality filter was applied. Only ~25% (approx. 260,000) of the generated RIRs were retained based on:
  • T60 Reverberation Time: Must be within ±20% of the reference distribution (excluding values > 1.8695s).
  • Direct-to-Reverberant Ratio (DRR): Must align with physical expectations for specific distances (excluding distances < 0.8m and > 7.1m).
  • Energy Decay & Reflections: Must match typical acoustic behavior and early reflection patterns of real-world measurements.
• Hyperparameter Optimization: An off-the-shelf algorithm was used to systematically search for optimal hyperparameters (Learning Rate: $10^{-5}$ to $10^{-3}$; Epochs: 5–50) to minimize test error on the provided Treble IRs.
• Model Training: The filtered, high-quality augmented RIRs were used to fine-tune a state-of-the-art SDE model.

3. Key Contributions

1. Conditioned Generative Modeling: Demonstrated the effectiveness of modifying FastRIR to condition exclusively on source-receiver positioning for RIR augmentation.
2. Quality Filtering Framework: Established a robust methodology for filtering synthetic acoustic data based on physical acoustic metrics (T60, DRR, energy decay), ensuring the generated data distribution matches real-world constraints.
3. Empirical Validation: Provided evidence that well-designed data augmentation significantly improves SDE accuracy, particularly in medium-to-long distance ranges.
4. Simulation-Specific Tuning: Showed that maintaining separate generative models for different simulation environments (Treble vs. GWA) yields better results than a unified approach.

4. Results

The proposed approach yielded significant improvements over the baseline SDE model:

• Performance Analysis:
  • The model showed superior alignment with ground truth, particularly for GWA rooms (MAE 0.6m) compared to Treble rooms (MAE 0.69m).
  • Distance Dependency: The model performed consistently well for distances > 1 meter (errors typically ~0.5m).
  • Limitations: Accuracy decreased significantly for distances < 1 meter. This is attributed to a lack of training examples in the close-proximity range and the unique acoustic phenomena at very short distances that are difficult to model without specialized data.
• Specialized Models: Training separate SDE systems exclusively on Treble or GWA augmented data further improved performance (10% MAE reduction for Treble, 5% for GWA) compared to the unified model.
• Visual Evidence: Time-domain comparisons confirmed that the fine-tuned models generated distinct reflection patterns consistent with their specific training datasets.

5. Significance and Future Work

This work demonstrates that generative data augmentation is a viable and powerful strategy for overcoming data scarcity in acoustic modeling. By filtering synthetic data to ensure physical realism, the authors successfully reduced estimation errors by more than half in challenging scenarios.

Future Directions:

• Combining multiple generative approaches to increase diversity.
• Developing specialized models specifically for extreme close-range estimation (< 1m).
• Extending these techniques to more diverse, uncontrolled acoustic environments beyond the simulated room settings used in the challenge.