Repository files navigation

Sum-product networks (SPNs) with Gaussian leaves are used as speaker models for automatic speaker identification (ASI)[1]. An example of an SPN with univariate Gaussian leaves is shown in Figure 1. Marginalisation and bounded marginalisation, as proposed byCooket al., are used to significantly increase the robustness of the SPN speaker models to noise. To identify the reliable spectral components for marginalisation, ana priori SNR estimator is used.

Figure 1: SPN speaker model with univariate Gaussian leaves.

1. git clone https://github.com/anicolson/SPN-ASI.git
2. cd SPN-ASI
3. virtualenv --system-site-packages -p python3 ~/venv/SPN-ASI
4. source ~/venv/SPN-Spk-Rec/bin/activate
5. pip install -r requirements.txt

TIMIT corpus for clean speech:

The clean speech of the speakers included in the TIMIT corpus are used to train the ASI system. The path to the TIMIT corpus is set inconfig.sh. If there is an issue with SPHERE .wav files, please use the followingscript.

Noisy speech anda priori SNR estimates:

The noisy speech anda priori SNR estimates used for testing the ASI system can be obtained from:http://dx.doi.org/10.21227/nbec-q510. Please see theDeep Xi repository if you require outputs from ana priori SNR estimator. The paths to the noisy speech anda priori SNR estimates are set inconfig.sh.

Training:

./run.sh TRAIN 1

Identification:

./run.sh IDENTIFICATION 1 MARG 1 BOUNDS 1

MARG is the flag for the marginalisation, andBOUNDS specifies whether bounds are to be used with marginalisation.

Please cite the following:

[1] Nicolson, A., & Paliwal, K. K. (2020). Sum-Product Networks for Robust Automatic Speaker Identification. Proc. Interspeech 2020.