Search Results (5)

Parameter-efficient fine-tuning (PEFT) methods like Low-Rank Adaptation (LoRA) are widely used to adapt large pre-trained models under limited resources, yet they often underperform full fine-tuning in low-resource automatic speech recognition (ASR). This gap stems partly from initialization strategies that ignore speech signals’ inherent spectral-phase structure. Unlike SVD/QR-based approaches (PiSSA, OLoRA) that construct mathematically optimal but signal-agnostic subspaces, we propose SPaRLoRA (Spectral-Phase Residual LoRA), which leverages Discrete Fourier Transform (DFT) bases to create speech-aware low-rank adapters. SPaRLoRA explicitly incorporates both magnitude and phase information by concatenating real and imaginary parts of DFT basis vectors, and applies residual correction to focus learning exclusively on components unexplained by the spectral subspace. Evaluated on a 200-h Sichuan dialect ASR benchmark, SPaRLoRA achieves a 2.1% relative character error rate reduction over standard LoRA, outperforming variants including DoRA, PiSSA, and OLoRA. Ablation studies confirm the individual and complementary benefits of spectral basis, phase awareness, and residual correction. Our work demonstrates that signal-structure-aware initialization significantly enhances parameter-efficient fine-tuning for low-resource ASR without architectural changes or added inference cost. Full article

In this study, we propose the CVMFCC-DR (Complex-Valued Mel-Frequency Cepstral Coefficients Dimensionality Reduction) algorithm as an efficient method for reducing the dimensionality of speech signals. By utilizing the complex-valued MFCC technique, which considers both real and imaginary components, our algorithm enables dimensionality reduction without information loss while decreasing computational costs. The efficacy of the proposed algorithm is validated through experiments which demonstrate its effectiveness in building a speech recognition model using a complex-valued neural network. Additionally, a complex-valued softmax interpretation method for complex numbers is introduced. The experimental results indicate that the approach yields enhanced performance compared to traditional MFCC-based techniques, thereby highlighting its potential in the field of speech recognition. Full article

Recent speech enhancement studies have mostly focused on completely separating noise from human voices. Due to the lack of specific structures for harmonic fitting in previous studies and the limitations of the traditional convolutional receptive field, there is an inevitable decline in the auditory quality of the enhanced speech, leading to a decrease in the performance of subsequent tasks such as speech recognition and speaker identification. To address these problems, this paper proposes a Harmonic Repair Large Frame enhancement model, called HRLF-Net, that uses a harmonic repair network for denoising, followed by a real-imaginary dual branch structure for restoration. This approach fully utilizes the harmonic overtones to match the original harmonic distribution of speech. In the subsequent branch process, it restores the speech to specifically optimize its auditory quality to the human ear. Experiments show that under HRLF-Net, the intelligibility and quality of speech are significantly improved, and harmonic information is effectively restored. Full article

In recent years, a lot of researchers’ attentions were concentrating on imaginary speech understanding, decoding, and even recognition. Speech is a complex mechanism, which involves multiple brain areas in the process of production, planning, and precise control of a large number of muscles and articulation involved in the actual utterance. This paper proposes an intelligent imaginary speech recognition system of eleven different utterances, seven phonemes, and four words from the Kara One database. We showed, during our research, that the feature space of the cross-covariance in frequency domain offers a better perspective of the imaginary speech by computing LDA for 2D representation of the feature space, in comparison to cross-covariance in the time domain and the raw signals without any processing. In the classification stage, we used a CNNLSTM neural network and obtained a performance of 43% accuracy for all eleven different utterances. The developed system was meant to be a subject’s shared system. We also showed that, using the channels corresponding to the anatomical structures of the brain involved in speech production, i.e., Broca area, primary motor cortex, and secondary motor cortex, 93% of information is preserved, obtaining 40% accuracy by using 29 electrodes out of the initial 62. Full article

Speech is a complex mechanism allowing us to communicate our needs, desires and thoughts. In some cases of neural dysfunctions, this ability is highly affected, which makes everyday life activities that require communication a challenge. This paper studies different parameters of an intelligent imaginary speech recognition system to obtain the best performance according to the developed method that can be applied to a low-cost system with limited resources. In developing the system, we used signals from the Kara One database containing recordings acquired for seven phonemes and four words. We used in the feature extraction stage a method based on covariance in the frequency domain that performed better compared to the other time-domain methods. Further, we observed the system performance when using different window lengths for the input signal (0.25 s, 0.5 s and 1 s) to highlight the importance of the short-term analysis of the signals for imaginary speech. The final goal being the development of a low-cost system, we studied several architectures of convolutional neural networks (CNN) and showed that a more complex architecture does not necessarily lead to better results. Our study was conducted on eight different subjects, and it is meant to be a subject’s shared system. The best performance reported in this paper is up to 37% accuracy for all 11 different phonemes and words when using cross-covariance computed over the signal spectrum of a 0.25 s window and a CNN containing two convolutional layers with 64 and 128 filters connected to a dense layer with 64 neurons. The final system qualifies as a low-cost system using limited resources for decision-making and having a running time of 1.8 ms tested on an AMD Ryzen 7 4800HS CPU. Full article