In recent years, deep learning models have demonstrated remarkable success in various domains, such as computer vision, natural language processing, and speech recognition. However, the generalization capabilities of these models can be negatively impacted by the limitations of their feature fusion techniques. This paper introduces an innovative approach, Adaptive Feature Fusion (AFF), to enhance the generalization of deep learning models by dynamically adapting the fusion process of feature representations. The proposed AFF framework is designed to incorporate fusion layers into existing deep learning architectures, enabling seamless integration and improved performance. By leveraging a combination of data-driven and model-based fusion strategies, AFF is able to adaptively fuse features based on the underlying data characteristics and model requirements. This paper presents a detailed description of the AFF framework, including the design and implementation of fusion layers for various architectures. Extensive experiments are conducted on multiple benchmark datasets, with the results demonstrating the superiority of the AFF approach in comparison to traditional feature fusion techniques. The analysis showcases the effectiveness of AFF in enhancing generalization capabilities, leading to improved performance across different tasks and applications. Finally, the paper discusses various real-world use cases where AFF can be employed, providing insights into its practical applicability. The conclusion highlights the potential for future research directions, including the exploration of advanced fusion strategies and the extension of AFF to other machine learning paradigms.
The attention mechanism has largely improved the performance of end-to-end speech recognition systems. However, the underlying behaviours of attention is not yet clearer. In this study, we use decision trees to explain how the attention mechanism impact itself in speech recognition. The results indicate that attention levels are largely impacted by their previous states rather than the encoder and decoder patterns. Additionally, the default attention mechanism seems to put more weights on closer states, but behaves poorly on modelling long-term dependencies of attention states.
Speaker recognition is a biometric modality that utilizes the speaker's speech segments to recognize the identity, determining whether the test speaker belongs to one of the enrolled speakers. In order to improve the robustness of the i-vector framework on cross-channel conditions and explore the nova method for applying deep learning to speaker recognition, the Stacked Auto-encoders are used to get the abstract extraction of the i-vector instead of applying PLDA. After pre-processing and feature extraction, the speaker and channel-independent speeches are employed for UBM training. The UBM is then used to extract the i-vector of the enrollment and test speech. Unlike the traditional i-vector framework, which uses linear discriminant analysis (LDA) to reduce dimension and increase the discrimination between speaker subspaces, this research use stacked auto-encoders to reconstruct the i-vector with lower dimension and different classifiers can be chosen to achieve final classification. The experimental results show that the proposed method achieves better performance than the state-of-the-art method.
Self-supervised learning (SSL) has allowed substantial progress in Automatic Speech Recognition (ASR) performance in low-resource settings. In this context, it has been demonstrated that larger self-supervised feature extractors are crucial for achieving lower downstream ASR error rates. Thus, better performance might be sanctioned with longer inferences. This article explores different approaches that may be deployed during the fine-tuning to reduce the computations needed in the SSL encoder, leading to faster inferences. We adapt a number of existing techniques to common ASR settings and benchmark them, displaying performance drops and gains in inference times. Interestingly, we found that given enough downstream data, a simple downsampling of the input sequences outperforms the other methods with both low performance drops and high computational savings, reducing computations by 61.3% with an WER increase of only 0.81. Finally, we analyze the robustness of the comparison to changes in dataset conditions, revealing sensitivity to dataset size.
Intent classification is a fundamental task in the spoken language understanding field that has recently gained the attention of the scientific community, mainly because of the feasibility of approaching it with end-to-end neural models. In this way, avoiding using intermediate steps, i.e. automatic speech recognition, is possible, thus the propagation of errors due to background noise, spontaneous speech, speaking styles of users, etc. Towards the development of solutions applicable in real scenarios, it is interesting to investigate how environmental noise and related noise reduction techniques to address the intent classification task with end-to-end neural models. In this paper, we experiment with a noisy version of the fluent speech command data set, combining the intent classifier with a time-domain speech enhancement solution based on Wave-U-Net and considering different training strategies. Experimental results reveal that, for this task, the use of speech enhancement greatly improves the classification accuracy in noisy conditions, in particular when the classification model is trained on enhanced signals.
This study addresses robust automatic speech recognition (ASR) by introducing a Conformer-based acoustic model. The proposed model builds on a state-of-the-art recognition system using a bi-directional long short-term memory (BLSTM) model with utterance-wise dropout and iterative speaker adaptation, but employs a Conformer encoder instead of the BLSTM network. The Conformer encoder uses a convolution-augmented attention mechanism for acoustic modeling. The proposed system is evaluated on the monaural ASR task of the CHiME-4 corpus. Coupled with utterance-wise normalization and speaker adaptation, our model achieves $6.25\%$ word error rate, which outperforms the previous best system by $8.4\%$ relatively. In addition, the proposed Conformer-based model is $18.3\%$ smaller in model size and reduces total training time by $79.6\%$.
This work explores the use of constant-Q transform based modulation spectral features (CQT-MSF) for speech emotion recognition (SER). The human perception and analysis of sound comprise of two important cognitive parts: early auditory analysis and cortex-based processing. The early auditory analysis considers spectrogram-based representation whereas cortex-based analysis includes extraction of temporal modulations from the spectrogram. This temporal modulation representation of spectrogram is called modulation spectral feature (MSF). As the constant-Q transform (CQT) provides higher resolution at emotion salient low-frequency regions of speech, we find that CQT-based spectrogram, together with its temporal modulations, provides a representation enriched with emotion-specific information. We argue that CQT-MSF when used with a 2-dimensional convolutional network can provide a time-shift invariant and deformation insensitive representation for SER. Our results show that CQT-MSF outperforms standard mel-scale based spectrogram and its modulation features on two popular SER databases, Berlin EmoDB and RAVDESS. We also show that our proposed feature outperforms the shift and deformation invariant scattering transform coefficients, hence, showing the importance of joint hand-crafted and self-learned feature extraction instead of reliance on complete hand-crafted features. Finally, we perform Grad-CAM analysis to visually inspect the contribution of constant-Q modulation features over SER.
We combine recent advancements in end-to-end speech recognition to non-autoregressive automatic speech recognition. We push the limits of non-autoregressive state-of-the-art results for multiple datasets: LibriSpeech, Fisher+Switchboard and Wall Street Journal. Key to our recipe, we leverage CTC on giant Conformer neural network architectures with SpecAugment and wav2vec2 pre-training. We achieve 1.8%/3.6% WER on LibriSpeech test/test-other sets, 5.1%/9.8% WER on Switchboard, and 3.4% on the Wall Street Journal, all without a language model.
Speech enhancement (SE) aims to suppress the additive noise from a noisy speech signal to improve the speech's perceptual quality and intelligibility. However, the over-suppression phenomenon in the enhanced speech might degrade the performance of downstream automatic speech recognition (ASR) task due to the missing latent information. To alleviate such problem, we propose an interactive feature fusion network (IFF-Net) for noise-robust speech recognition to learn complementary information from the enhanced feature and original noisy feature. Experimental results show that the proposed method achieves absolute word error rate (WER) reduction of 4.1% over the best baseline on RATS Channel-A corpus. Our further analysis indicates that the proposed IFF-Net can complement some missing information in the over-suppressed enhanced feature.
With the development of automatic speech recognition (ASR) and text-to-speech (TTS) technology, high-quality voice conversion (VC) can be achieved by extracting source content information and target speaker information to reconstruct waveforms. However, current methods still require improvement in terms of inference speed. In this study, we propose a lightweight VITS-based VC model that uses the HuBERT-Soft model to extract content information features without speaker information. Through subjective and objective experiments on synthesized speech, the proposed model demonstrates competitive results in terms of naturalness and similarity. Importantly, unlike the original VITS model, we use the inverse short-time Fourier transform (iSTFT) to replace the most computationally expensive part. Experimental results show that our model can generate samples at over 5000 kHz on the 3090 GPU and over 250 kHz on the i9-10900K CPU, achieving competitive speed for the same hardware configuration.