Introduction
In modern speech enhancement, multi-microphone (multi-channel) arrays are essential for capturing spatial information to filter out noise and reverberation. However, standard Deep Neural Networks (DNNs) struggle with "parameter explosion" as the number of microphones increases. Researchers have proposed a novel approach using Tensor-Train Networks (TTN) to perform tensor-to-vector regression, successfully shrinking model sizes by over 80% while simultaneously improving speech quality.
The Parameter Explosion Problem
Standard DNN architectures typically handle multi-channel audio by concatenating features from every microphone into a single, massive input vector. This approach presents two significant challenges:
- Input Size Explosion: Adding microphones directly increases the input vector dimension.
- Hidden Layer Expansion: To maintain expressive power, hidden layers must often be wider than the input dimension.
These factors lead to models with tens of millions of parameters, making them impractical for real-time applications or deployment on resource-constrained edge devices.
The Tensor-Train Network Approach
To address these limitations, the authors propose moving from traditional vector-to-vector regression to a tensor-based framework. Instead of flattening multi-channel data, the TTN treats the input from $B$ microphones as a 4th-order tensor, which more naturally aligns with the multi-dimensional structure (time, frequency, and spatial channel) of audio data.
As shown in the architecture above, the TTN utilizes Tensor-Train Decomposition. Rather than storing a massive weight matrix $W$, the network represents weights as a sequence of small, low-rank "core tensors." Unlike post-processing compression techniques like Singular Value Decomposition (SVD), TTNs can be initialized randomly and trained from scratch, ensuring the model learns an efficient representation from the start.
Experimental Results
The models were evaluated using clean speech from the WSJ0 corpus and 90 types of additive noise from the OSU-100-noise dataset. Room environments were simulated using the Image-Source Method (ISM) with an 8-channel circular microphone array.
The results demonstrate that TTNs consistently outperform standard DNNs and SVD-compressed models across various configurations:
- 8-Channel Performance: A standard DNN required 68 million parameters to reach a PESQ (Perceptual Evaluation of Speech Quality) score of 3.06. In contrast, a TTN achieved a superior PESQ of 3.12 using only 20 million parameters.
- Extreme Compression: Even when the 8-channel model was shrunk to just 5 million parameters (a 93% reduction compared to the full DNN), the TTN maintained a PESQ of 3.06, matching the performance of the much larger baseline.
- Efficiency Gains: When compared at similar parameter budgets, the TTN consistently provided an absolute improvement of approximately 0.1 in PESQ scores over standard architectures.
Conclusion
By aligning the model architecture with the inherent tensor structure of multi-channel audio, Tensor-Train Networks break the link between input dimensionality and parameter count. This allows for high expressive power without the storage and computational penalties of traditional DNNs. Future work will explore applying Tensor-Train decomposition to other architectures, such as Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs), to further optimize audio processing pipelines.
Summary updated