TLDR: A new unsupervised method for speech dereverberation trains a neural network using only reverberant speech. It works by estimating a dry signal, then reverberating it with a classical model, and comparing the result to the original reverberant input. This “reverberation auto-supervision” outperforms existing unsupervised methods by achieving more consistent improvements across various speech quality metrics, especially in data-scarce scenarios.
Reverberation, the phenomenon where sound waves reflect off surfaces in a room, significantly degrades the clarity and intelligibility of recorded speech. This is a common problem in many acoustic environments, from conference calls to professional recordings. Traditionally, systems designed to remove this reverberation, known as dereverberation systems, have relied on a large amount of paired data – recordings of both “dry” (unreverberated) speech and its reverberated counterpart. However, obtaining such perfectly matched dry and reverberant speech pairs is often a challenging and resource-intensive task.
A new research paper, “Déréverbération non-supervisée de la parole par modèle hybride” (Unsupervised Speech Dereverberation by Hybrid Model), introduces an innovative unsupervised training strategy that tackles this data scarcity problem head-on. Authored by Louis Bahrman, Mathieu Fontaine, and Gaël Richard from LTCI, Télécom Paris, Institut Polytechnique de Paris, this work proposes a method that can improve speech dereverberation systems using only reverberant speech signals. This is a significant step forward, as reverberant speech is much easier to collect than paired dry/reverberant data.
The Hybrid Approach: Learning Without Direct Supervision
The core idea behind this new approach, termed “reverberation auto-supervision,” is to train a deep neural network (DNN) to predict a clean, dry speech signal from a reverberant input. What makes it unsupervised is how the system learns: instead of comparing the predicted dry signal to an actual dry reference, it uses a clever feedback loop. The estimated dry speech signal is fed into a classical reverberation model, which then synthesizes a new reverberant signal. The system then compares this newly synthesized reverberant signal to the original reverberant input. The goal is for these two reverberant signals to match as closely as possible, effectively teaching the DNN to produce a dry signal that, when reverberated, recreates the original input.
This hybrid model leverages limited acoustic information, such as the reverberation time (RT60), which describes how long it takes for sound to decay in a room. This parameter is estimated from the reverberant signal itself, further reducing the need for external, labeled data. The process involves a non-parametric Room Impulse Response (RIS) synthesizer and a convolutional model, which do not require training, keeping the complexity focused on the dereverberation DNN.
Robust Performance Across Metrics
The researchers conducted extensive experiments, comparing their proposed method against state-of-the-art unsupervised dereverberation techniques, specifically MetricGAN-U. The results, evaluated using objective metrics like Scale-Invariant Signal to Distortion Ratio (SISDR), Extended Short-Time Objective Intelligibility (ESTOI), and Wide-Band Perceptual Evaluation of Speech Quality (WB-PESQ), demonstrated the superiority of the new approach.
While MetricGAN-U excelled in optimizing its specific target metric (SRMR), it often did so at the expense of other crucial quality indicators, sometimes even degrading the speech quality compared to the original reverberant input. In contrast, the reverberation auto-supervision method achieved more consistent and balanced improvements across all evaluated metrics, indicating a more generalized and robust dereverberation performance. Even when the reverberation time (RT60) was blindly estimated rather than provided as an oracle, the performance remained remarkably close to the weakly supervised variant, highlighting the method’s resilience to estimation errors.
Also Read:
- Robust Emotion Recognition in Speech: Disentangling Features and Aligning Embeddings
- Unpacking Sequence Model Behavior Through Eigenvalue Analysis
Future Implications
This unsupervised hybrid model represents a significant advancement in speech dereverberation, particularly for scenarios where obtaining paired dry/reverberant data is impractical. By relying solely on reverberant speech and limited acoustic parameters, it opens up new possibilities for developing effective dereverberation solutions in diverse real-world applications. The authors plan to extend this work to generative auto-supervised approaches, further enhancing the model’s capabilities. For more technical details, you can refer to the full research paper available at arXiv.
