Deep Xi (where the Greek letter 'xi' or ξ is pronounced /zaɪ/) is a deep learning approach to a priori SNR estimation that was proposed in [1]. It can be used by minimum mean-square error (MMSE) approaches to speech enhancement like the MMSE short-time spectral amplitude (MMSE-STSA) estimator, the MMSE log-spectral amplitude (MMSE-LSA) estimator, and the Wiener filter (WF) approach. It can also be used to estimate the ideal ratio mask (IRM) and the ideal binary mask (IBM). Deep Xi can be used as a front-end for robust ASR, as shown in Figure 1. Deep Xi is implemented in TensorFlow.

Figure 1: Deep Xi used as a front-end for robust ASR. The back-end (Deep Speech) is available here. The noisy speech magnitude spectrogram, as shown in (a), is a mixture of clean speech with voice babble noise at an SNR level of -5 dB, and is the input to Deep Xi. Deep Xi estimates the a priori SNR, as shown in (b). The a priori SNR estimate is used to compute an MMSE approach gain function, which is multiplied elementwise with the noisy speech magnitude spectrum to produce the clean speech magnitude spectrum estimate, as shown in (c). MFCCs are computed from the estimated clean speech magnitude spectrogram, producing the estimated clean speech cepstrogram, as shown in (d). The back-end system, Deep Speech, computes the hypothesis transcript, from the estimated clean speech cepstrogram, as shown in (e).

Figure 2: ResLSTM a priori SNR estimator.

Figure 3: ResBLSTM a priori SNR estimator.

The scripts for each of the following a priori SNR estimators can be found in the ./ver directory:

• c2.7a is a TCN (temporal convolutional network) that has 2 million parameters.
• c1.13a is a ResLSTM (residual long short-term memory network) with 10.8 million parameters, as shown in Figure 2.
• n1.9a is a ResBLSTM (residual bidirectional long short-term memory network) with 21.3 million parameters, as shown in Figure 3.

'c' and 'n' indicate if the system is causal or non-causal, respectively. 'c1', 'n1', and 'c2' indicate if a ResLSTM, ResBLSTM, or a TCN network is used, respectively.

• c2.7a can be found in the ./model directory.
• c1.13a can be downloaded from Dropbox here, or the Nihao cloud service here.
• n1.9a can be downloaded from Dropbox here, or the Nihao cloud service here.

Note: The trained models for c1.13a and n1.9a are too large to be stored on github.

It is recommended to use a virtual environment for installation.

Prerequisites:

• TensorFlow r1.11 (installed in a virtual environment). Will be updated to r2.0 in the near future.
• Python3
• MATLAB (only required for .mat output files)

To install:

1. git clone https://github.com/anicolson/DeepXi.git
2. pip install -r requirements.txt

Inference:

cd ver/c2/7/a
python3 deepxi.py --test 1 --out_type y --gain mmse-lsa --gpu 0

y for --out_type specifies enhanced speech .wav output. mmse-lsa specifies the used gain function (others include mmse-stsa, wf, irm, ibm, srwf, cwf).

Training:

cd ver/c2/7/a
python3 deepxi.py --train 1 --verbose 1 --gpu 0

Other options can be found in the deepxi.py script.

The .wav files used for training are single-channel, with a sampling frequency of 16 kHz.

The following speech datasets were used:

• The train-clean-100 set from Librispeech corpus, which can be found here.
• The CSTR VCTK corpus, which can be found here.
• The si and sx training sets from the TIMIT corpus, which can be found here (not open source).

The following noise datasets were used:

• The QUT-NOISE dataset, which can be found here.
• The Nonspeech dataset, which can be found here.
• The Environemental Background Noise dataset, which can be found here.
• The noise set from the MUSAN corpus, which can be found here.
• Multiple packs from the FreeSound website, which can be found here

[1] A. Nicolson and K. K. Paliwal, "Deep Learning For Minimum Mean-Square Error Approaches to Speech Enhancement", Submitted with revisions to Speech Communication.