[lines=2]Speech enhancement technology is essential to improve the intelligibility and quality of noisy speech signal[speech_enhancement]. Classcial speech enhancments techniques include spectral subtraction, Wiener filtering
, miniimum mean-square- error(MMSE) estimator and the optimally modified log-spectral amplitude speech estimator[2001Speech]
. These convetional methods based on time-frequency domain achieve relatively good performance in stationary-noise environment, whereas they are not robust enough in tackling non-stationary noises in most scenes.
Over past few years, deep nerual networks (DNNs) have significantly evaluated the performance of speech enhancement[2017Supervised]. Existing DNN apporaches provide better results compared to classic techniques. In [DBLP:conf/interspeech/MaasLOVNN12][2017A], recurrent nerual network (RNNs) is advanced to model temporal features. Auto-encoder is employed for speech enhancement in [DBLP:conf/interspeech/LuTMH13]
, and the first speech enhancemen benchmark ultilizes DNN as non-linear regression function[2015A]. A convolutional recurrent neural network (CRN) is used to extract long-context information[inproceedings222]
. Speech enhancement techniques also take example by image synthesis that using Generative adversarial network(GAN) architecture to reconstruct target speech signal[2017SEGAN, 2019MetricGAN, Lin2019, 2019Sergan, DBLP:conf/interspeech/LinNWMSW20]. These DNN supervised-based speech enhancements methods are outperform than classic algorithms.
Recently, U-net structures have achieved significant success and have overran their performance than basic DNN architectures in various machine learning tasks including medical diagnostics[2015U], semantic segmentation[2017SegNet], singing source separation[2017Sing] and others. Motivated by this sucess, speech enhancement has explorerd U-net structures in both raw waveform[2018WaveUnet, 2018ImprovedSE, 2018SpeechDerevb, 2017SEGAN] and time-frequency features[2018TimeF]. In [A2019Demucs], Wave-U-Net employed GLU activation in encoder and decoder, as well as bidirectional LSTM in-between. DCCRN[2020DCCRN]takes advantages of u-net and deep complex network [DCN] for denoising. The combination between attention-unit and U-net boost the performance of speech enhancment one step further. Self-attention is an efficient context information aggregation mechanism that operates on the input sequence itself and that can be utilizedfor any task that has a sequential input and output. Attention wave-u-net [2019Attention] outperforms all other published speech enhancement approaches on the Voice BankCorpus (VCTK) dataset.
Mask-based target, which describes the time-frequency relationships between clean speech and background noise,is utilized to train the network. Typically, conventinal masks consist of ideal binary mask(IBM)[IBM], ideal ratio mask(IRM)[IRM] and spectral magnitude mask(SMM)[SMM] only consider the magnitude between clean waveform and mixture audio. Subsequently, phase is taken into account, where phase-sensitive mask(PSM)[PSM] is the first method to show the feasibility of phase information, and complex ratio mask (CRM)[CRM] is announced that it can reconstruct speech perectly by enhancing both real and imaginary components of the division of clean speech and mixture speech spectrogram simultaneously. Soon afterwards, CRN[CRN] used one encoder and two decoders for complex spectral mapping(CSM) to coincidently estimate the real and imaginary spectrogram of mixture speech. It isworth noting that CRM and CSM possess the full informationof a speech signal so that they can achieve the best oracle speech enhancement performance in theory.
In this paper, we propose a network called CARN incorporating effitive components of u-net, attention mechanism, skip connection, LSTMs in between encoder-decoder and CRM in time-frequency domain. We also Investigated conbining gate convolution network with CARN model into a model,named GCARN. We show that, based on the speech quality metrics (PESQ, etc.) on the 2020 DNS challenge[reddy2020interspeech], and the dataset released by Valentini[valentini2017noisy]. the attention-lstm-unet utilized crm in time-frequency domain achieve significant imporvement results outperforming others published speech enhancement methods on these datasets.
2 The CARN Model
The CRN model was firstly proposed by Ke Tan[tanCRN],and was further investigated with complex spectral mapping[CRN] mechanism and gate convolution[GCRN]. According to the above research, we investigated a new model named CARN,which was combined CRN with attention mechanism.In the CARN model.we used attention based skip connections from encoder layers to decoder layers. Furthermore,we constrast our proposed model with the gate convolution based CRN model,as well as several previous works.
2.1 The CARN Architecture
The encoder and decoder both consist of 6 Conv2d blocks with PReLU activation function,aimming at extracting high-demensional features from the input features as well as reducing the resolution.Between encoder and decoder,we used two LSTM layers to study temporal features. The model is illustrated as Figure1
.We took spectrogram feature as input. The LSTM layers hidden size is 512,The T-F kernel size is 3, stride is, for each Conv2d or ConvTranspose2d layer. Each Conv2d or ConvTranspose2d layer is followed by a batchnorm layer. A linear layer is embedded after last ConvTranspose2d layer to map the complex ratio mask(CRM) from the ouput features. At last,CRM multiply with the input stft spectrogram to get clean stft spectrogram,refering to (4) and (5). All the activation function is PReLU.
2.2 Attention Mechanism
Different from conventional CRN architecture, self-attention masks are applied to multiply with the output of encoder by skip-connection. The output of attention layer is concatenated to the last decoder output for the next decoder input.
The AttentionBlock[2019Attention] is described in Figure 2. is the output of encoder architecture and is the output of LSTM layers or decoder covolution layers. Additional two 2-d convolutions, with kernel size 3, output channels twice of input channels, refering as and , which are used to mapping and to high-dimensional space feature,are used to model the attention mechanism. The high-dimensional space feature demension is twice of the demension of ; The high-dimensional space feature layer output can be discriped as (1).
where and present layer of encoder and decoder, respectively. is sigmoid fuction. The output of self-attention block,
where presents another 2-d concolution layer.
2.3 Training targets
CARN estimates CRM and is optimized by signal approximation (SA).Given the complex-vauled STFT spectrogram of clean speech S and noisy speech Y,CRM can be defined as
where and refer as the real and imaginary parts of the noisy complex spectrogram,respectively. and refer as the real and imaginary parts of the clean complex spectrogram, similarly.Let and denote the real and imaginary parts of the estimating denoise audio complex spectrogram,respectively. and denote the real and imaginary parts of CRM,then
2.4 Loss function
We train the model with loss function as
where and denote estimating denoise audio and clean audio respectively. ,is the power-compressd STFTs. This loss function consists of spectrogram MSE (mean square error) and power-compressd STFTs MSE.
In our experiments, we evaluate the proposed models on two datasets.
3.1.1 Dataset 1: Noisy speech database
The first dataset111https://datashare.ed.ac.uk/handle/10283/2791, is released by Valentini et. al.[valentini2017noisy], widely used by speech enhancement research, and it generalizes on various types of noise for different speakers. This dataset includes clean and noisy audio data at 48kHz sampling frequency, which requires downsampling to 16kHz for training and testing. The clean sets are recording of sentences, sourced from various text passages, and thirty English-speakers, including male and female with various accents, are selected from the Voice Bank corpus[veaux2013voice]. 28 and 2 speakers are assigned to the training and test sets, respectively. The test set consists of 20 different noise conditions, 5 type of noise sourced from the DEMAND database ,yielding 824 test items, with approximately 20 different sentences in each condition per test speaker.[valentini2017noisy]
3.1.2 Dataset 2: DNS 2020
The second dataset is based on the data provided by the Interspeech 2020 DNS Challenge dataset [reddy2020interspeech]. The DNS Challenge dataset consists of 180-hour noise set which includes 150 classes and 65,000 noise clips, and over 500 hours clean speech, which includes audio clips from 2150 speakers. The clean speech dataset is derived from the public audiobooks dataset called Librivox. The noise clips were selected from Audioset and Freesound. We randomly selected 24000 speakers with all noise clips to created 200-hour noisy train set,with singal-noise-ratio ranging from 0dB to 40dB. Each selected speaker’s audio clips are concatenated to 30 seconds,while mixing various noise clips. We estimated the proposed model with DNS-Challenge no-blind test dataset and blind test dataset.Both datasets consists of synthesis dataset and real-recordings.
3.2 Training setup and baselines
We trained the proposed model on both dataset using Adam optimizer annealed with warmup schedule, learning rate of 1e-3 and 64 mini-batchsize. The model was selected by early stopping when the was smaller than 0.05. The STFT window length and hop size are modulated to 32 millisecond and 16 millisecond respectively. Hanning window and 512 FFT length are applied. We stacked the real and imaginary parts of FFT spectrogram together as input.We compared the proposed model with several models,described as followed.
CRN: A causal complex spectral mapping Convolution Recurrent Network[CRN], is evaluated on the first dataset as contrast. This model has same structure as the proposed model CARN except attention mechanism. It consists of six convolution layers encoder and six symmetrical layers decoder with two lstm layers between thems.
GCRN: A causal complex spectral mapping Gate Convolution Recurrent Network[GCRN], is evaluated on the first dataset as constrast.This model has same structure as the CRN above,except gate convolution layers.
DCCRN-E: The model proposed by [2020DCCRN]
, which is based on the CRN architecture but complex convolution layers, has achieved the 1st place in the 2020 Deep Noise Suppression Challenge’s Real-Time Track. We compare this model under two datasets using officially opened source code222 https://github.com/huyanxin/DeepComplexCRN.
CARN: The proposed model,well described at section 2, is training and testing on both datasets,respectively.
GCARN: Same structure as the CARN, except gate convolution layers, is a contrast model.
3.3 Evaluation:Objective Metrics and Result
We evaluated wide-band Perceptual Evaluation of Speech Quality(PESQ) -ITU-T P.862.2 [PESQ],and the composite CSIG,CBAK,and COVL [compositeMetrics] and STOI,on both test dataset.These metrics evaluated my model performance in some respects. But these metrics require reference clean speech,and cannot work at real recordings. The blind test dataset in DNS-Challenge 2020 consist of real recordings and synthetic noisy audios. We evaluated our proposed models on this dataset with an another metric,called DNSMOS.
DNSMOS metric was proposed by Chandan K A Reddy,etc.[DNSMOS] recently.The DNSMOS metric is a non-intrusive objective speech quality metric for wide band scenario,and more reliable than other widely used objective metrics such as SDR and POLDA, and does not require reference clean speech.So it can work on real recordings.
In Table 1, we listed the results over the first dataset of several methods ,such as Wiener,SEGAN[2017SEGAN],U-Net[2017Sing], WaveNet[2018WaveUnet],CRN,GCRN ,DCCRN-E and our proposed models CARN and GCARN. CRN and GCRN models were better than Wiener,SEGAN,U-Net and WaveNet,except CBAK scores, and were not as good as DCCRN-E at PESQ. We improved CRN and GCRN by replacing the directly connection with attention between encoder layers and decoder layers,which we called CARN and GCARN models respectively,achieving significant improvement in all these metrics.CARN improved CRN on PESQ,CBAK,COVL,CSIG with 12.2%,10.7%,11.7%,10.8% respectively,and GCARN improved GCRN on these metrics with 19.1%,6.8%,17.5%,15.1% respectively. CARN and GCARN were the best models among them. When comparing CRN and CARN with GCRN and GCARN,the results demonstrated the gate convolution may be unecessary in this issue.
In Table 2,we trained and evaluated CARN and GCARN models on DNS-Challenge dataset.We compared our proposed models with RNNoise,the DNS-Challenge baseline[reddy2020interspeech],DCCRN-E,and poCoNet [Poconet] which took 1st place in the 2020 Deep Noise Suppression Challenge’s Non-Real-Time Track. GCARN tied with CARN,and they achieved better scores in all these metrics,compared with these models.CARN outperformed DCCRN-E at PESQ,CBAK,COVL,CSIG with 15.5%,5.2%,13.2%,10.4% relatively,and outperformed poCoNet at these metrics with 6.2%,20.7%,5.3%,3.7% relatively.
In Table 3,we estimated our proposed models and DCCRN-E over the DNS-Challenge 2020 blind test dataset, with the DNSMOS metric.CARN and GCARN can achieve significant imporvement speech quality compared with DCCRN-E.CARN scored the best in all non-reverb ,reverb and real-recordings.
The experiments over the two datasets shows that attention mechanism can significantly improve CRN architecture performance when compared with directly connection. A reasonable explanation is that attention mechanism filters some noise features which are connected from encoder layer to decoder layer.Our proposed model CARN outperformed the place 1st model in both real time and non-real time track of the DNS Challenge 2020 at many metrics.