Efficient Streaming Speech Quality Prediction with Spiking Neural Networks

This repository contains the code used for training and evaluating the convolutional spiking neural networks (CSNNs) and baseline ANNs presented in our Interspeech 2025 paper.

This is a monorepo containing two installable Python packages:

• sqp-snn: Training and evaluation of SNNs using stork
• sqp-ann: Training and evaluation of baseline ANNs, adapted from our Interspeech 2024 SQP code

For adapting NeuroBench to stork, we adapted code from fmi-basel/neural-decoding-RSNN.

Installation

We used Python 3.10.12 for the development of the code. We recommend installing each of the two packages independently with the following steps:

1. Create and activate a new virtual environment (recommended)

   python -m venv sqp-snn-env
   source sqp-snn-env/bin/activate

2. In spiking-sqp, navigate to package directory

   cd sqp-snn

3. Install requirements

   pip install -r requirements.txt

4. Install package:

   pip install -e .

5. Repeat the steps above independently for sqp-ann

Dataset

We used the Interspeech 2020 DNS Challenge dataset augmented by denoising with Dynamic NSNet2.

A link to our augmented dataset will be added here soon.

Generate Training Data

You can generate non-augmented training data by following these steps:

1. Audio data: Run the download and single-process data generation scripts from the Interspeech 2020 DNS Challenge. Change the following parameters in noisyspeech_synthesizer.cfg to match our training data prior to augmentation:

   • audio_length: 5
   • total_hours: 50
   • snr_lower: -15
   • snr_upper: 25

2. Labels: Compute PESQ speech quality labels by inserting your dataset path in compute_labels.py and then running the script

   python compute_labels.py

3. Repeat #2 for the test set

Prepare for training

1. .env file: Create a .env file in sqp-ann/ by copying the provided example

   cp sqp-ann/example.env sqp-ann/.env

2. Paths: Update sqp-snn/src/sqp_snn/project_config.py and sqp-ann/.env with the correct paths for audio data and metrics files (labels)

3. Reader IDs: Generate the readers.pkl file used for reader-based validation split by running:

   python generate_readers_file.py

Run experiments

CLI note: sqp-snn uses argparse, while sqp-ann uses Hydra. See examples below for the respective command-line argument formats.

Train and evaluate CSNNs

Training time: Training a CSNN took up to 9 hours on an NVIDIA A4000 depending on the number of epochs to early stopping on convergence.

In sqp-snn/:

• r-CSNN: Train recurrent CSNN

   python train.py --model recurrent 

• ff-CSNN: Train feedforward CSNN

   python train.py --model feedforward

• Sparsity experiments: Train sequence of models with different degrees of spike-rate regularization yielding the results in Fig. 4 of the paper

   bash sparsity_experiments_recurrent.sh

   bash sparsity_experiments_feedforward.sh

• Benchmarking: Measure EFLOPs and activation sparsity for a trained CSNN logged to W&B with NeuroBench by running the notebook benchmark_snn.ipynb

• CLI: See src/sqp_snn/args.py for all possible command-line arguments

• Logging: You can activate W&B logging by setting --logger to wandb and entering your W&B information in src/sqp_snn/args.py

Train and evaluate baseline ANNs

In sqp-ann/:

• DNSMOS: Train a 2D CNN with the architecture of DNSMOS (this trains a CNN from scratch on PESQ-labeled data – it does not imply computing the metric called DNSMOS)

   python train.py model=dnsmos

  • Benchmarking: Benchmark a trained DNSMOS model with NeuroBench by running the notebook benchmark_dnsmos.ipynb

• Quality-Net: Train a causal version of Quality-Net based on UniLSTM

   python train.py model=qualitynet_realtime

  • Benchmarking: LSTMs lack activation sparsity and were not supported by NeuroBench at the time of our experiments — instead, multiply–accumulate operations are counted with torchinfo by default when you launch training

• CLI: See sqp-ann/config/train.yaml for possible command-line arguments

• Logging: Enable W&B logging by uncommenting - logger_wandb in config/train.yaml