Low SNR Speech Research Articles

SNR-Progressive Model with Harmonic Compensation for Low-SNR Speech Enhancement

SNR-Progressive Model with Harmonic Compensation for Low-SNR Speech Enhancement

CST: Complex Sparse Transformer for Low-SNR Speech Enhancement.

Speech enhancement tasks for audio with a low SNR are challenging. Existing speech enhancement methods are mainly designed for high SNR audio, and they usually use RNNs to model audio sequence features, which causes the model to be unable to learn long-distance dependencies, thus limiting its performance in low-SNR speech enhancement tasks. We design a complex transformer module with sparse attention to overcome this problem. Different from the traditional transformer model, this model is extended to effectively model complex domain sequences, using the sparse attention mask balance model's attention to long-distance and nearby relations, introducing the pre-layer positional embedding module to enhance the model's perception of position information, adding the channel attention module to enable the model to dynamically adjust the weight distribution between channels according to the input audio. The experimental results show that, in the low-SNR speech enhancement tests, our models have noticeable performance improvements in speech quality and intelligibility, respectively.

Low SNR speech enhancement with DNN based phase estimation

In low Signal-to-Noise Ratio environment phase information is one of the important factor and therefore this article consider the importance of clean phase in single channel speech enhancement technique. The proposed method uses Deep Neural Network based regression model to estimate clean phase and clean amplitude for speech reconstruction. Experiments are conducted over five different noises such as factory, restaurant, car, airport and babble at different levels and result are evaluated using objective quality measures like Perceptual Evaluation of Speech Quality, Weighted Spectral Slope, Cepstrum Distance, frequency weighted segmented Signal-to-Noise Ratio and Log Likelihood Ratio. The overall quality of speech improved for factory noise by $$12\%$$ , restaurant noise by $$8\%$$ , car noise by $$13\%$$ , airport noise by $$10\%$$ and babble noise by $$14\%$$ respectively.

Intelligibility for Binaural Speech with Discarded Low-SNR Speech Components.

Speech intelligibility in multitalker settings improves when the target speaker is spatially separated from the interfering speakers. A factor that may contribute to this improvement is the improved detectability of target-speech components due to binaural interaction in analogy to the Binaural Masking Level Difference (BMLD). This would allow listeners to hear target speech components within specific time-frequency intervals that have a negative SNR, similar to the improvement in the detectability of a tone in noise when these contain disparate interaural difference cues. To investigate whether these negative-SNR target-speech components indeed contribute to speech intelligibility, a stimulus manipulation was performed where all target components were removed when local SNRs were smaller than a certain criterion value. It can be expected that for sufficiently high criterion values target speech components will be removed that do contribute to speech intelligibility. For spatially separated speakers, assuming that a BMLD-like detection advantage contributes to intelligibility, degradation in intelligibility is expected already at criterion values below 0 dB SNR. However, for collocated speakers it is expected that higher criterion values can be applied without impairing speech intelligibility. Results show that degradation of intelligibility for separated speakers is only seen for criterion values of 0 dB and above, indicating a negligible contribution of a BMLD-like detection advantage in multitalker settings. These results show that the spatial benefit is related to a spatial separation of speech components at positive local SNRs rather than to a BMLD-like detection improvement for speech components at negative local SNRs.

Toeplitz Robust Noisy Speech Endpoint Detection

In this paper, under the conditions of low SNR speech endpoint detection, a feature based on the maximum value of Toeplitz Noise endpoint detection methods. Terms of the method of spectrum from the corresponding sequences with a symmetric Toeplitz matrix constructed using the maximum eigenvalue of the matrix information on the voice signal for dual endpoint detection threshold. New algorithm has been tested to effectively distinguish between speech and noise, low-noise in different environmental conditions has good robustness. With the recent recursive signal analysis methods, the accuracy is higher. The algorithm to calculate the cost of a small, real good, simple and easy to implement.