This repository accompanies the paper:
M. Hidalgo-Araya et al., "A Probabilistic Generative Model for Spectral Speech Enhancement", 2025.
A comprehensive evaluation framework for virtual hearing aids using the VOICEBANK_DEMAND dataset with warped filter bank (WFB) preprocessing.
This revision of the repository reports two complementary speech-quality metric families side by side for transparency:
- The DNSMOS P.835 non-intrusive metrics, reported as
DSIG,DBAK,DOVRL(D-prefix so they stay visually parallel to the composite metrics and never get confused with them), used in the initial submission, and - The Hu & Loizou (2008) composite metrics (CSIG, CBAK, COVL), which are the P.835-aligned intrusive metrics requested during the review cycle.
Both families are produced from the same evaluation run when run_evaluation.jl is invoked with --composite; earlier DNSMOS-only runs remain byte-identical when the flag is omitted.
If you want to verify the paper's results end-to-end, run:
julia --project=. scripts/run_paper_results.jlThat one script resamples the dataset (Stage 1), generates the WFB-processed reference (Stage 2), evaluates all four paper configurations with --composite --checkpoint-interval 50 (Stage 3), and regenerates the benchmark tables in this README (Stage 4). It is idempotent: re-running after completion only refreshes the tables, and re-running after an interrupted evaluation resumes from the last checkpoint. Expect a few hours on a laptop for the full 824-file testset × 4 systems × 7 metrics.
Two follow-up commands produce the paper's LaTeX tables and parameter-evolution figure directly from the runs above:
julia --project=. scripts/generate_latex_tables.jl # -> tables/*.tex
julia --project=. scripts/plot_parameter_evolution.jl # -> figures/parameter_evolution_bus_7p5dB_band13.*
julia --project=. scripts/generate_results_md.jl # -> RESULTS.md (markdown summary)The one-time prerequisites (clone, Julia + Python dependencies, dataset download) are listed in the Installation and Step 1 sections below.
This repository provides the complete implementation and evaluation framework for the spectral speech enhancement model presented in the paper. It includes:
- Implementation: Full codebase for the Warped-Frequency Filter Bank (WFB) front-end and Speech Enhancement Model (SEM) backend
- Evaluation Pipeline: Automated evaluation on the VOICEBANK_DEMAND dataset with comprehensive metrics (PESQ, DNSMOS, and optionally the Hu & Loizou CSIG/CBAK/COVL composite metrics)
- Reproducibility: All configurations and scripts needed to reproduce the results reported in the paper, including the uniform-filter-bank (uFB) ablation
- Benchmark Comparisons: Automated generation of comparison tables
This repository provides a complete pipeline for:
- Dataset Preparation: Download, resample, and preprocess VOICEBANK_DEMAND dataset
- WFB Preprocessing: Create warped filter bank processed dataset for consistent evaluation
- Evaluation: Run evaluations for baseline and hearing aid algorithms using
run_evaluation.jl - Results Analysis: Generate summary tables and metrics organized by SNR and environment
- Benchmark Results: Automatically generate and update benchmark comparison tables in the README
The latest benchmark results comparing different hearing aid algorithms are automatically generated and displayed in the Benchmark Results section below. To update these results with the latest evaluation runs, simply run:
julia scripts/update_readme_benchmark.jlThis script automatically:
- Finds the latest runs for each hearing aid (excluding Baseline_clean)
- Generates comprehensive comparison tables for:
- Overall summary across all metrics
- Performance by SNR level (2.5, 7.5, 12.5, 17.5 dB)
- Performance by environment and SNR (bus, cafe, living, office, psquare)
- Updates the README with the latest results and configuration details
- Julia 1.11+: Required for all functionality
- Python 3.7+: Required for metrics evaluation (PESQ, DNSMOS)
- Git: For cloning and submodule management
- Clone the repository with submodules:
git clone --recursive https://github.com/biaslab/Publication_Spectral_Subtraction.git
cd Publication_Spectral_SubtractionThe --recursive flag is required because the repository pulls microsoft/DNS-Challenge as the python_modules/DNSMOS/ submodule. DNSMOS ships as a standalone upstream project; vendoring it as a submodule pins the exact revision used for the paper's results and puts the ONNX models at the path the Julia wrapper expects (python_modules/DNSMOS/DNSMOS/DNSMOS/{sig_bak_ovr,model_v8}.onnx), so no manual path plumbing is needed.
If you already cloned without --recursive, initialize the submodule after the fact:
git submodule update --init --recursive --depth=1--depth=1 keeps the DNS-Challenge checkout shallow (~3 MB of ONNX weights plus the reference Python; the full upstream repo is ~275 MB and most of it is unused training data). If you plan to work with the upstream DNS-Challenge data, drop --depth=1.
- Install Julia dependencies:
using Pkg
Pkg.activate(".")
Pkg.instantiate()- Install Python dependencies for metrics:
python install_python_deps.pyThis installs pesq, the bundled dnsmos_wrapper, and pysepm (used by the optional Hu & Loizou CSIG/CBAK/COVL composite metrics). pysepm is installed directly from its GitHub archive because it is not on PyPI; if the install fails for any reason, the composite metrics are disabled at runtime and PESQ+DNSMOS continue to work as before.
Download the VOICEBANK_DEMAND dataset from the official source:
- Visit the official dataset page: https://datashare.ed.ac.uk/handle/10283/2791
- Download the dataset files
- Extract and place them in the following structure:
databases/VOICEBANK_DEMAND/
├── data/
│ ├── clean_testset_wav/ # Clean audio files
│ └── noisy_testset_wav/ # Noisy audio files
├── logfiles/
│ └── log_testset.txt # SNR information
└── testset_txt/ # Text transcriptions
Resample the VOICEBANK_DEMAND dataset to 16 kHz using Julia:
using HADatasets
# Create dataset instance pointing to the database directory
dataset = HADatasets.VOICEBANKDEMANDDataset("databases/VOICEBANK_DEMAND")
# Resample with default settings (16kHz, 1.0s minimum duration)
HADatasets.resample_data(dataset)This creates:
databases/VOICEBANK_DEMAND_resampled/
├── clean_testset_wav/ # Resampled clean files
├── noisy_testset_wav/ # Resampled noisy files
└── logfiles/
└── log_testset_resampled.txt # Updated log file
Note: The resampled dataset preserves the same directory structure as the original, with all audio files resampled to 16 kHz.
Why WFB preprocessing is needed:
The hearing aid processing pipeline uses a Warped Filter Bank (WFB) that warps the frequency domain of the audio. Since PESQ is sensitive to changes in the data or missing samples, we need to ensure consistent preprocessing for fair evaluation.
The WFB preprocessing:
- Processes all audio through the BaselineHearingAid (which has unity gains, so the audio is unaltered except for the WFB warping)
- Creates a preprocessed dataset where all files have been through the same WFB pipeline
- Ensures that when we evaluate hearing aids, we compare against a consistent WFB-processed clean reference
Create the WFB dataset:
julia scripts/convert_to_wfb.jlOr test with a limited number of samples first:
julia scripts/convert_to_wfb.jl --num-samples=10This script:
- Loads the BaselineHearingAid configuration
- Processes all clean and noisy files from
VOICEBANK_DEMAND_resampledthrough the WFB - Creates
VOICEBANK_DEMAND_resampled_wfb/with the same directory structure:databases/VOICEBANK_DEMAND_resampled_wfb/ ├── clean_testset_wav/ # WFB-processed clean files ├── noisy_testset_wav/ # WFB-processed noisy files ├── logfiles/ # Copied logfiles
Note: If the WFB dataset already exists, the script will detect it and skip processing with the following messages:
[Info: WFB dataset already exists and appears to be processed
[Info: Skipping conversion - dataset already processed
All evaluations, including baselines and hearing aid algorithms, are run using the run_evaluation.jl script:
Before evaluating hearing aids, establish baseline scores for comparison:
Baseline Best (Clean vs Clean) - Upper bound performance:
julia scripts/run_evaluation.jl configurations/baseline_clean/baseline_clean.tomlBaseline Unprocessed (Clean vs Noisy) - Lower bound performance:
julia scripts/run_evaluation.jl configurations/baseline_noise/baseline_noise.tomlEvaluate each hearing aid algorithm on the WFB-processed dataset:
# Evaluate SEM Hearing Aid with the warped filter bank front-end (apcoefficient = 0.5)
julia scripts/run_evaluation.jl configurations/SEMHearingAid/SEMHearingAid.toml
# Evaluate SEM Hearing Aid with the uniform filter bank front-end (apcoefficient = 0.0)
julia scripts/run_evaluation.jl configurations/SEMHearingAid_uFB/SEMHearingAid_uFB.tomlThe two configurations above correspond to the WFB ablation reported in the OJ-SP revision (SectionVI "Structural Analysis of the WFB" and AppendixE). They differ only in the filter-bank warping coefficient.
# Test with a single file first
julia scripts/run_evaluation.jl configurations/SEMHearingAid/SEMHearingAid.toml --single-file p257_001.wav
# Limit number of samples for testing
julia scripts/run_evaluation.jl configurations/SEMHearingAid/SEMHearingAid.toml --num-samples 50
# Custom checkpoint interval (save every N files)
julia scripts/run_evaluation.jl configurations/SEMHearingAid/SEMHearingAid.toml --checkpoint-interval 20
# Save processed output audio files
julia scripts/run_evaluation.jl configurations/SEMHearingAid/SEMHearingAid.toml --save-output
# Also compute the Hu & Loizou (2008) CSIG/CBAK/COVL composite metrics
# (requires pysepm; installed by install_python_deps.py)
julia scripts/run_evaluation.jl configurations/SEMHearingAid/SEMHearingAid.toml --composite--composite adds three columns (CSIG, CBAK, COVL) to results.csv, overall_summary.csv, summary_by_snr.csv, and summary_by_environment_snr.csv. When the flag is omitted, the output files are byte-identical to the pre-composite pipeline, so runs without --composite remain directly comparable to earlier releases (tag v1.1.1).
Results are organized in timestamped directories:
results/VOICEBANK_DEMAND/
├── BaselineHearingAid/
│ └── run_<timestamp>/
│ ├── BaselineHearingAid.toml
│ └── table/
│ ├── results.csv # Complete results for all files
│ ├── overall_summary.csv # Overall average scores
│ ├── summary_by_snr.csv # Average scores by SNR level
│ ├── summary_by_environment_snr.csv # Average scores by environment and SNR
│ └── checkpoint_*.csv # Optional checkpoint files (if --checkpoint-interval used)
│ └── run_<timestamp>/
│ └── ...
└── SEMHearingAid/
└── run_<timestamp>/
└── ...
Each evaluation computes the following metrics:
- PESQ (Perceptual Evaluation of Speech Quality): 1-5 scale, higher is better
- DSIG (Signal Quality from DNSMOS): 1-5 scale, higher is better
- DBAK (Background Quality from DNSMOS): 1-5 scale, higher is better
- DOVRL (Overall Quality from DNSMOS): 1-5 scale, higher is better
When invoked with --composite, the evaluation additionally computes three Hu & Loizou (2008) composite metrics:
- CSIG (Composite Signal): 1-5 scale, higher is better; predicts MOS for speech distortion.
- CBAK (Composite Background): 1-5 scale, higher is better; predicts MOS for background intrusiveness.
- COVL (Composite Overall): 1-5 scale, higher is better; predicts the overall MOS.
CSIG/CBAK/COVL are linear regressions of PESQ, LLR, WSS, and segSNR tuned against ITU-T P.835 subjective ratings. They are reference-based (require the clean signal) and use the pysepm implementation of the Loizou reference code. If pysepm is not available at runtime, --composite raises a clear error and the non-composite pipeline is unaffected.
The evaluation automatically generates:
overall_summary.csv: Overall average scores across all conditionssummary_by_snr.csv: Average scores for each SNR level (2.5, 7.5, 12.5, 17.5 dB)summary_by_environment_snr.csv: Average scores per environment per SNR levelresults.csv: Complete results for all individual files
- Automatic checkpoints: Saved every N files (default: 10, configurable) - checkpoint files are created when using
--checkpoint-intervaloption - Resume capability: If evaluation is interrupted, checkpoints can be merged manually
- Final results: All results are saved to
results.csvin the table directory
After running evaluations for multiple hearing aids, you can automatically generate and update benchmark comparison tables in the README:
julia scripts/update_readme_benchmark.jlThis script:
- Finds the latest runs for each hearing aid (excluding Baseline_clean)
- Generates comprehensive comparison tables showing:
- Overall summary across all metrics (PESQ, DSIG, DBAK, DOVRL, and optionally CSIG, CBAK, COVL)
- Performance breakdown by SNR level (2.5, 7.5, 12.5, 17.5 dB)
- Performance breakdown by environment and SNR (bus, cafe, living, office, psquare)
- Updates the README with the latest results and configuration details
The benchmark results are displayed in the Benchmark Results section below.
This repository uses comprehensive speech quality assessment metrics to evaluate hearing aid algorithms. All metrics are computed using the HADatasets module, which provides standardized implementations of ITU-T and IEEE/ACM standards.
- Type: Intrusive (requires reference signal)
- Scale: 1-5 (higher is better)
- Standard: ITU-T P.862.2
- Use Case: Overall speech quality assessment
- Description: PESQ is a perceptual metric that predicts the subjective quality of speech as perceived by human listeners. It compares the processed/enhanced audio to the clean reference signal and provides a score that correlates with Mean Opinion Score (MOS) ratings.
Important Note: PESQ is sensitive to changes in the data or missing samples. This is why the evaluation pipeline uses WFB-processed clean audio as the reference, ensuring that both the processed output and reference have undergone the same WFB preprocessing for fair comparison.
- Type: Non-intrusive (no reference required)
- Scale: 1-5 (higher is better)
- Standard: Microsoft DNS Challenge P.835
- Use Case: Noise suppression quality assessment
- Description: DNSMOS is a deep learning-based metric that predicts subjective quality scores without requiring a clean reference signal. It follows the ITU-T P.835 subjective test framework to measure three key quality dimensions.
P.835 Dimensions:
-
DOVRL (DNSMOS Overall Quality): Overall audio quality assessment
- Measures the overall perceived quality of the processed audio
- Combines both speech and background noise quality perceptions
-
DSIG (DNSMOS Signal Quality): Speech quality assessment
- Focuses specifically on the quality of the speech signal
- Measures how natural and clear the speech sounds
-
DBAK (DNSMOS Background Quality): Background noise quality assessment
- Evaluates the quality of the background/noise component
- Measures how well noise is suppressed while preserving speech
The D-prefix is deliberate: it keeps these DNSMOS columns visually parallel to the CSIG/CBAK/COVL composite metrics so no one reading a table confuses the two families.
- Type: Intrusive (requires reference signal)
- Scale: 1-5 (higher is better)
- Reference: Hu, Y. and Loizou, P. C. (2008). "Evaluation of Objective Quality Measures for Speech Enhancement." IEEE Trans. Audio Speech Lang. Process. 16(1), 229–238.
- Use Case: Predict subjective P.835 MOS ratings directly from the time-domain clean/enhanced pair without the DNSMOS neural net.
- Description: CSIG, CBAK, and COVL are linear regressions of PESQ, LLR (log-likelihood ratio), WSS (weighted spectral slope), and segmental SNR against subjective ratings collected under the ITU-T P.835 protocol. We use the
pysepmimplementation of Loizou's reference MATLAB code. - Availability: Computed only when
run_evaluation.jlis invoked with--composite; requirespysepm(installed automatically byinstall_python_deps.py).
The combination of PESQ, DNSMOS, and (optionally) the Hu & Loizou composite metrics provides a comprehensive evaluation:
- PESQ provides an intrusive reference-based assessment, giving a direct comparison to the clean signal.
- DNSMOS provides a non-intrusive assessment that doesn't require a reference, making it useful for real-world scenarios where clean references may not be available. Its three P.835 dimensions (
DOVRL,DSIG,DBAK) summarize overall, speech, and background quality from a deep acoustic model. - CSIG/CBAK/COVL provide a second, classical P.835-aligned readout derived from well-established time/frequency features, and are reported alongside DNSMOS for transparency and cross-check.
This evaluation framework adopts the ITU-T P.835 subjective test framework to measure speech enhancement quality across multiple dimensions, enabling comprehensive assessment of hearing aid algorithms for monaural speech enhancement tasks.
Spectral_Subtraction/
├── databases/
│ ├── VOICEBANK_DEMAND/ # Original dataset (downloaded)
│ ├── VOICEBANK_DEMAND_resampled/ # Resampled dataset (16 kHz)
│ └── VOICEBANK_DEMAND_resampled_wfb/ # WFB-processed dataset
├── configurations/
│ ├── baseline_clean/
│ ├── baseline_noise/
│ ├── BaselineHearingAid/
│ ├── SEMHearingAid/ # Paper algorithm (warped FB, apcoefficient=0.5)
│ └── SEMHearingAid_uFB/ # WFB ablation (uniform FB, apcoefficient=0.0)
├── results/
│ └── VOICEBANK_DEMAND/ # Evaluation results
├── scripts/
│ ├── convert_to_wfb.jl # WFB conversion script
│ ├── run_evaluation.jl # Per-config evaluation script
│ ├── run_paper_results.jl # One-command reproduction orchestrator
│ ├── generate_latex_tables.jl # Populates the paper's LaTeX tables from the latest runs
│ ├── plot_parameter_evolution.jl # Regenerates the parameter-evolution figure (s/n, ξ, w̃)
│ └── update_readme_benchmark.jl # Benchmark results update script
├── src/
│ ├── Experiments.jl # Main evaluation module
│ ├── HADatasets/ # Dataset loaders and metrics (PESQ, DNSMOS, CSIG/CBAK/COVL)
│ ├── HASoundProcessing/ # SEM factor graph + inference rules
│ └── VirtualHearingAid/ # WFB front-end and hearing-aid backends
└── python_modules/ # PyCall-side wrappers (dnsmos_wrapper, composite_wrapper)
The SEM follows the model introduced in the paper:
The Speech Enhancement Model (SEM) uses a probabilistic generative model for Bayesian inference of speech and noise characteristics, enabling adaptive spectral enhancement.
The WFB front-end provides perceptually-aligned frequency warping for consistent evaluation:
The input signal passes through a cascade of first-order all-pass filters, producing warped delay-line signals. A time-domain FIR structure with weights generates the output, while the warped signals are provided to the Spectral Enhancement Model for inference and synthesis.
- Input: WFB-processed noisy audio (
VOICEBANK_DEMAND_resampled_wfb/noisy_testset_wav/) - Processing: Pass through hearing aid algorithm
- Reference: WFB-processed clean audio (
VOICEBANK_DEMAND_resampled_wfb/clean_testset_wav/) - Metrics: Compare processed output to WFB-processed clean reference
- BaselineHearingAid: Unity gain processing (no noise reduction, WFB only)
- SEMHearingAid: Speech Enhancement Model (Bayesian inference)
If you just want to regenerate every table in the paper, including the WFB ablation and the composite metrics added in the revision, run:
julia --project=. scripts/run_paper_results.jlThis single orchestrator:
- Resamples VoiceBank+DEMAND to 16 kHz (skipped if already done).
- Generates the WFB-processed reference dataset (skipped if already done).
- Evaluates the four configurations used by the paper tables with
--composite --checkpoint-interval 50, so every per-file row and every summary CSV carries the seven metrics (PESQ, DSIG, DBAK, DOVRL, CSIG, CBAK, COVL):configurations/baseline_clean/baseline_clean.toml(upper bound)configurations/baseline_noise/baseline_noise.toml(unprocessed lower bound)configurations/SEMHearingAid/SEMHearingAid.toml(paper algorithm, warped FB)configurations/SEMHearingAid_uFB/SEMHearingAid_uFB.toml(WFB ablation, uniform FB)
- Calls
scripts/update_readme_benchmark.jlto refresh the markdown tables at the bottom of this README.
The orchestrator is idempotent: re-running it after a clean pass only regenerates the README tables; re-running after an interrupted evaluation resumes from the last checkpoint.
Before running it once, you still need to:
- Clone the repository with
git clone --recursive …(so thepython_modules/DNSMOS/submodule is initialized; see Installation above). - Instantiate the Julia project:
julia --project=. -e 'using Pkg; Pkg.instantiate()'. - Install the Python dependencies:
python install_python_deps.py. - Download the raw VoiceBank+DEMAND corpus into
databases/VOICEBANK_DEMAND/as described in Step 1.
If you prefer to run each stage yourself (for debugging, or to run only a subset):
-
Prepare the
VOICEBANK_DEMAND_resampled_wfbdataset by following Steps 1 and 2 in this README. -
Run the hearing-aid configurations, passing
--compositeso that each run reports both the DNSMOS metrics (DSIG,DBAK,DOVRL) and the Hu & Loizou composite metrics (CSIG,CBAK,COVL):# Main paper algorithm (warped filter bank, apcoefficient = 0.5) julia scripts/run_evaluation.jl configurations/SEMHearingAid/SEMHearingAid.toml --composite --checkpoint-interval 50 # WFB ablation (uniform filter bank, apcoefficient = 0.0) julia scripts/run_evaluation.jl configurations/SEMHearingAid_uFB/SEMHearingAid_uFB.toml --composite --checkpoint-interval 50
-
Update the README tables:
julia scripts/update_readme_benchmark.jl
-
Generate the paper's LaTeX tables directly from the same runs:
julia --project=. scripts/generate_latex_tables.jl
This populates three files under
tables/:tables/tab_comparison_with_params.tex— one-row-per-system comparison with PESQ / CSIG / CBAK / COVL means (and a±line of per-file sample standard deviations for the SEM row). Populates\label{tab:comparison_with_params}.tables/tab_metrics_quadrants.tex— the full 4-metric × 5-environment × 4-SNR × 3-system ablation, every cell formatted as$\mathrm{mean} \pm \mathrm{std}$over the files assigned to that (environment, SNR) cell. Populates\label{tab:metrics-quadrants}.tables/tab_per_env_delta.tex— per-environment improvement summary, averaged across the four input SNRs:Δ_{U→W}= SEM (WFB) − Unprocessed, andΔ_{u→W}= SEM (WFB) − SEM (uFB) (WFB ablation). The best positive improvement per column is rendered in\mathbf{}. Populates\label{tab:per-env-delta}.
The script reads the latest
run_*/table/results.csvfor each ofbaseline_noise,SEMHearingAid_uFB,SEMHearingAid. Drop\input{tables/<filename>}in the paper body to use them. -
The results used in the paper correspond to the runs in:
results/VOICEBANK_DEMAND/<Device>/run_<timestamp>/Each run directory contains a single
results.csvwith all seven metrics (PESQ, DSIG, DBAK, DOVRL, CSIG, CBAK, COVL) plus three summary CSVs broken down by SNR and by (environment, SNR), ready to paste into the paper.
To add a new hearing aid algorithm:
-
Implement the backend in
src/VirtualHearingAid/(create a new<Name>Backendtype). -
Create a configuration file in
configurations/<NewHearingAid>/<NewHearingAid>.toml:[parameters.hearingaid]withtype = "<NewHearingAid>"[parameters.frontend]WFB parameters (nbands, fs, etc.)[parameters.backend.*]for algorithm-specific parameters
-
Run evaluation:
julia scripts/run_evaluation.jl configurations/<NewHearingAid>/<NewHearingAid>.toml
-
Update the benchmark tables:
julia scripts/update_readme_benchmark.jl
See existing configurations in configurations/ for examples of the TOML structure.
- Tested on: macOS / Linux, Julia 1.11+, Python 3.7+
- GPU: Not required. All models are CPU-friendly
- Storage: ~2 GB for the resampled dataset, ~4 GB for the WFB-processed dataset
- Missing files: Ensure the dataset is downloaded and extracted correctly
- Resampling errors: Check that audio files are valid WAV files
- WFB conversion fails: Verify BaselineHearingAid configuration exists
- Memory errors: Use
--num-samplesto process in smaller batches - Checkpoint errors: Manually merge existing checkpoints if needed
- Metrics errors: Ensure Python dependencies are installed (
python install_python_deps.py) ImportError: No module named 'dnsmos_local': Thepython_modules/DNSMOS/submodule was not initialized. Rungit submodule update --init --recursive --depth=1from the repository root.- CSIG/CBAK/COVL missing from results:
--compositewas not passed torun_evaluation.jl, orpysepmfailed to install. Re-runpython install_python_deps.pyand confirmpysepmimports in Python.
The metrics evaluation functionality relies on Python integration and the following optional dependencies:
- PyCall: Python integration (for full metrics functionality)
- pesq: Python PESQ implementation (MIT License)
- dnsmos_wrapper: Custom wrapper for Microsoft DNSMOS (Creative Commons Attribution 4.0 International)
- pysepm: Python port of Loizou's composite-metrics MATLAB code, used for CSIG/CBAK/COVL (MIT-style licence; installed from the upstream GitHub archive)
These dependencies are automatically installed when running the Python installation script:
python install_python_deps.pyIf pysepm cannot be installed for some reason, PESQ and DNSMOS continue to work; only the optional --composite path is disabled.
Licensed under Creative Commons Attribution 4.0 International:
- Attribution Required: Must give appropriate credit to Microsoft
- Commercial Use: Allowed
- Modification: Allowed
- Distribution: Allowed
@inproceedings{reddy2022dnsmos,
title={DNSMOS P.835: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors},
author={Reddy, Chandan KA and Gopal, Vishak and Cutler, Ross},
booktitle={ICASSP 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
year={2022},
organization={IEEE}
}@inproceedings{dubey2023icassp,
title={ICASSP 2023 Deep Noise Suppression Challenge},
author={Dubey, Harishchandra and Aazami, Ashkan and Gopal, Vishak and Naderi, Babak and Braun, Sebastian and Cutler, Ross and Gamper, Hannes and Golestaneh, Mehrsa and Aichner, Robert},
booktitle={ICASSP},
year={2023}
}@misc{Valentini-Botinhao2017NoisySpeech,
author = {Valentini-Botinhao, Cassia},
title = {Noisy speech database for training speech enhancement algorithms and TTS models},
year = {2017},
howpublished = {Edinburgh DataShare},
doi = {10.7488/ds/2117},
url = {https://doi.org/10.7488/ds/2117}
}@article{hu2008evaluation,
title={Evaluation of Objective Quality Measures for Speech Enhancement},
author={Hu, Yi and Loizou, Philipos C.},
journal={IEEE Transactions on Audio, Speech, and Language Processing},
volume={16},
number={1},
pages={229--238},
year={2008},
publisher={IEEE}
}- ICASSP 2023 Deep Noise Suppression Challenge: Official challenge website and resources
- DNSMOS Implementation: Microsoft's DNS Challenge repository with DNSMOS implementation
- VoiceBank+Demand Dataset: Official dataset download page
This section presents benchmark results comparing different hearing aid algorithms on the VOICEBANK_DEMAND dataset.
| Device | PESQ (1-5) | DSIG — DNSMOS Signal (1-5) | DBAK — DNSMOS Background (1-5) | DOVRL — DNSMOS Overall (1-5) | CSIG — Composite Signal (1-5) | CBAK — Composite Background (1-5) | COVL — Composite Overall (1-5) |
|---|---|---|---|---|---|---|---|
| SEM | 2.167 | 3.286 | 3.467 | 2.777 | 3.444 | 2.614 | 2.78 |
| baseline_clean | 4.644 | 3.51 | 4.032 | 3.217 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.955 | 3.327 | 3.082 | 2.672 | 3.333 | 2.435 | 2.618 |
| SEM_uFB | 2.049 | 3.299 | 3.443 | 2.79 | 3.34 | 2.171 | 2.665 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.009 | 3.463 | 3.647 | 3.749 |
| baseline_clean | 4.034 | 4.036 | 4.028 | 4.031 |
| baseline_noise | 2.456 | 3.017 | 3.353 | 3.501 |
| SEM_uFB | 2.919 | 3.439 | 3.657 | 3.759 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.466 | 2.769 | 2.906 | 2.967 |
| baseline_clean | 3.211 | 3.22 | 3.218 | 3.221 |
| baseline_noise | 2.208 | 2.645 | 2.879 | 2.956 |
| SEM_uFB | 2.438 | 2.792 | 2.935 | 2.993 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.6 | 1.985 | 2.295 | 2.79 |
| baseline_clean | 4.644 | 4.644 | 4.644 | 4.644 |
| baseline_noise | 1.411 | 1.74 | 2.095 | 2.576 |
| SEM_uFB | 1.512 | 1.862 | 2.183 | 2.642 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.061 | 3.31 | 3.384 | 3.39 |
| baseline_clean | 3.502 | 3.511 | 3.513 | 3.515 |
| baseline_noise | 2.913 | 3.364 | 3.515 | 3.517 |
| SEM_uFB | 3.027 | 3.344 | 3.413 | 3.413 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.721 | 3.251 | 3.684 | 4.121 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 2.612 | 3.12 | 3.584 | 4.02 |
| SEM_uFB | 2.636 | 3.138 | 3.581 | 4.008 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.099 | 2.584 | 2.974 | 3.466 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.948 | 2.395 | 2.822 | 3.308 |
| SEM_uFB | 2.009 | 2.463 | 2.863 | 3.327 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.988 | 2.422 | 2.804 | 3.245 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.765 | 2.196 | 2.622 | 3.161 |
| SEM_uFB | 1.72 | 2.04 | 2.305 | 2.621 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.528 | 3.804 | 3.876 | 3.909 |
| baseline_clean | 4.05 | 4.054 | 4.011 | 4.007 |
| baseline_noise | 3.045 | 3.499 | 3.669 | 3.705 |
| SEM_uFB | 3.494 | 3.788 | 3.887 | 3.921 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.118 | 2.985 | 3.285 | 3.556 |
| baseline_clean | 4.027 | 4.049 | 4.033 | 4.065 |
| baseline_noise | 1.634 | 2.257 | 2.907 | 3.201 |
| SEM_uFB | 1.958 | 2.907 | 3.295 | 3.572 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.519 | 3.206 | 3.46 | 3.578 |
| baseline_clean | 4.027 | 4.032 | 4.012 | 3.993 |
| baseline_noise | 1.779 | 2.741 | 3.059 | 3.274 |
| SEM_uFB | 2.362 | 3.2 | 3.477 | 3.585 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.771 | 3.927 | 3.983 | 4.023 |
| baseline_clean | 4.047 | 4.023 | 4.048 | 4.054 |
| baseline_noise | 3.369 | 3.702 | 3.831 | 3.931 |
| SEM_uFB | 3.771 | 3.946 | 3.997 | 4.04 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.141 | 3.393 | 3.636 | 3.685 |
| baseline_clean | 4.019 | 4.021 | 4.038 | 4.036 |
| baseline_noise | 2.482 | 2.892 | 3.302 | 3.402 |
| SEM_uFB | 3.045 | 3.357 | 3.632 | 3.682 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.798 | 2.984 | 3.023 | 3.06 |
| baseline_clean | 3.216 | 3.249 | 3.189 | 3.194 |
| baseline_noise | 2.618 | 2.929 | 3.035 | 3.039 |
| SEM_uFB | 2.816 | 2.989 | 3.056 | 3.086 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.854 | 2.519 | 2.731 | 2.874 |
| baseline_clean | 3.214 | 3.247 | 3.231 | 3.283 |
| baseline_noise | 1.587 | 2.17 | 2.661 | 2.85 |
| SEM_uFB | 1.757 | 2.49 | 2.757 | 2.897 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.182 | 2.626 | 2.781 | 2.835 |
| baseline_clean | 3.204 | 3.21 | 3.205 | 3.158 |
| baseline_noise | 1.71 | 2.51 | 2.719 | 2.833 |
| SEM_uFB | 2.077 | 2.656 | 2.82 | 2.874 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.896 | 2.948 | 3.053 | 3.11 |
| baseline_clean | 3.224 | 3.175 | 3.228 | 3.238 |
| baseline_noise | 2.84 | 2.979 | 3.083 | 3.113 |
| SEM_uFB | 2.968 | 3.025 | 3.092 | 3.136 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.62 | 2.769 | 2.943 | 2.958 |
| baseline_clean | 3.197 | 3.215 | 3.237 | 3.231 |
| baseline_noise | 2.311 | 2.641 | 2.898 | 2.95 |
| SEM_uFB | 2.598 | 2.801 | 2.954 | 2.976 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.01 | 2.579 | 2.88 | 3.33 |
| baseline_clean | 4.644 | 4.644 | 4.644 | 4.644 |
| baseline_noise | 1.749 | 2.253 | 2.688 | 3.194 |
| SEM_uFB | 1.873 | 2.368 | 2.732 | 3.204 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.199 | 1.438 | 1.737 | 2.239 |
| baseline_clean | 4.644 | 4.644 | 4.644 | 4.644 |
| baseline_noise | 1.148 | 1.327 | 1.535 | 1.916 |
| SEM_uFB | 1.18 | 1.404 | 1.683 | 2.076 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.268 | 1.576 | 1.937 | 2.367 |
| baseline_clean | 4.644 | 4.644 | 4.644 | 4.644 |
| baseline_noise | 1.175 | 1.391 | 1.681 | 2.209 |
| SEM_uFB | 1.237 | 1.499 | 1.815 | 2.266 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.136 | 2.671 | 2.933 | 3.397 |
| baseline_clean | 4.644 | 4.644 | 4.644 | 4.644 |
| baseline_noise | 1.747 | 2.294 | 2.763 | 3.226 |
| SEM_uFB | 1.94 | 2.472 | 2.772 | 3.218 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.394 | 1.665 | 1.98 | 2.632 |
| baseline_clean | 4.644 | 4.644 | 4.644 | 4.644 |
| baseline_noise | 1.239 | 1.437 | 1.8 | 2.353 |
| SEM_uFB | 1.333 | 1.567 | 1.903 | 2.459 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.311 | 3.392 | 3.388 | 3.406 |
| baseline_clean | 3.5 | 3.532 | 3.49 | 3.497 |
| baseline_noise | 3.302 | 3.502 | 3.528 | 3.492 |
| SEM_uFB | 3.349 | 3.404 | 3.42 | 3.429 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.476 | 3.264 | 3.38 | 3.385 |
| baseline_clean | 3.505 | 3.532 | 3.525 | 3.566 |
| baseline_noise | 2.198 | 3.083 | 3.526 | 3.565 |
| SEM_uFB | 2.333 | 3.232 | 3.413 | 3.395 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.921 | 3.239 | 3.356 | 3.342 |
| baseline_clean | 3.494 | 3.502 | 3.507 | 3.464 |
| baseline_noise | 2.444 | 3.319 | 3.477 | 3.52 |
| SEM_uFB | 2.76 | 3.293 | 3.39 | 3.381 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.311 | 3.283 | 3.361 | 3.403 |
| baseline_clean | 3.514 | 3.474 | 3.514 | 3.524 |
| baseline_noise | 3.463 | 3.45 | 3.476 | 3.447 |
| SEM_uFB | 3.395 | 3.36 | 3.397 | 3.425 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.308 | 3.371 | 3.435 | 3.413 |
| baseline_clean | 3.496 | 3.513 | 3.532 | 3.523 |
| baseline_noise | 3.189 | 3.466 | 3.57 | 3.561 |
| SEM_uFB | 3.325 | 3.428 | 3.447 | 3.434 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.366 | 3.957 | 4.32 | 4.656 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 3.216 | 3.782 | 4.228 | 4.592 |
| SEM_uFB | 3.235 | 3.785 | 4.198 | 4.581 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.975 | 2.576 | 3.053 | 3.605 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.944 | 2.51 | 2.942 | 3.433 |
| SEM_uFB | 1.937 | 2.519 | 2.984 | 3.468 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.143 | 2.624 | 3.157 | 3.545 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 2.104 | 2.545 | 3.027 | 3.478 |
| SEM_uFB | 2.108 | 2.547 | 3.048 | 3.45 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 3.42 | 4.023 | 4.341 | 4.756 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 3.184 | 3.812 | 4.257 | 4.692 |
| SEM_uFB | 3.262 | 3.864 | 4.209 | 4.64 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.72 | 3.077 | 3.548 | 4.06 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 2.629 | 2.951 | 3.462 | 3.924 |
| SEM_uFB | 2.655 | 2.975 | 3.463 | 3.917 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.641 | 3.25 | 3.6 | 4.042 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 2.431 | 2.997 | 3.457 | 3.948 |
| SEM_uFB | 2.502 | 3.054 | 3.461 | 3.921 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.5 | 1.95 | 2.358 | 2.904 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.458 | 1.86 | 2.201 | 2.655 |
| SEM_uFB | 1.47 | 1.903 | 2.295 | 2.751 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.622 | 2.052 | 2.512 | 2.94 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.555 | 1.92 | 2.318 | 2.828 |
| SEM_uFB | 1.587 | 1.974 | 2.395 | 2.841 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.753 | 3.341 | 3.648 | 4.121 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 2.435 | 3.045 | 3.519 | 3.997 |
| SEM_uFB | 2.571 | 3.159 | 3.498 | 3.956 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.997 | 2.332 | 2.744 | 3.339 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.871 | 2.153 | 2.611 | 3.131 |
| SEM_uFB | 1.931 | 2.229 | 2.661 | 3.179 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.246 | 2.778 | 3.16 | 3.575 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.966 | 2.498 | 2.976 | 3.519 |
| SEM_uFB | 1.926 | 2.316 | 2.612 | 2.929 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.666 | 2.058 | 2.439 | 2.897 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.559 | 1.932 | 2.278 | 2.801 |
| SEM_uFB | 1.479 | 1.758 | 2.008 | 2.296 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.73 | 2.146 | 2.541 | 2.965 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.557 | 1.94 | 2.314 | 2.882 |
| SEM_uFB | 1.529 | 1.827 | 2.084 | 2.396 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 2.41 | 2.88 | 3.236 | 3.661 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 2.059 | 2.573 | 3.057 | 3.59 |
| SEM_uFB | 2.034 | 2.416 | 2.663 | 2.97 |
| Device | 2.5 dB | 7.5 dB | 12.5 dB | 17.5 dB |
|---|---|---|---|---|
| SEM | 1.899 | 2.248 | 2.639 | 3.135 |
| baseline_clean | 5.0 | 5.0 | 5.0 | 5.0 |
| baseline_noise | 1.691 | 2.038 | 2.479 | 3.023 |
| SEM_uFB | 1.64 | 1.886 | 2.156 | 2.521 |
The following configurations were used for each hearing aid:
[parameters.hearingaid]
name = "SEM Hearing Aid"
type = "SEMHearingAid"
processing_strategy = "BatchProcessingOffline"
[parameters.frontend]
name = "WFB"
type = "WFBFrontend"
nbands = 17
fs = 16000.0
spl_reference_db = 100.0
spl_power_estimate_lower_bound_db = 30.0
apcoefficient = 0.5
buffer_size_s = 0.0015
[parameters.backend.general]
name = "SEM"
type = "SEMBackend"
[parameters.backend.inference]
autostart = true
free_energy = false
iterations = 1
[parameters.backend.filters.time_constants90]
s = 5.0 # Speech time constant (ms)
n = 700.0 # Noise time constant (ms)
xnr = 20.0 # ξ time constant (ms)
[parameters.backend.priors.speech]
mean = 80.0
precision = 1.0
[parameters.backend.priors.noise]
mean = 80.0
precision = 1.0
[parameters.backend.gain]
threshold = 12.0 #(GMIN)
[parameters.backend.switch]
threshold = 2.0
[metadata]
author = "VirtualHearingAid"
date = "03-12-2025"
description = "SEM Hearing Aid configuration"
name = "SEM"
# Baseline Clean Configuration
# This configuration is used for baseline "best" evaluation (clean vs clean)
[metadata]
name = "Baseline Clean"
author = "VirtualHearingAid"
date = "2025-01-27"
description = "Baseline clean evaluation - compares WFB-processed clean audio to itself (best case scenario)"
# Baseline Noise Configuration
# This configuration is used for baseline "worst" evaluation (clean vs noisy)
[metadata]
name = "Baseline Noise"
author = "VirtualHearingAid"
date = "2025-01-27"
description = "Baseline noise evaluation - compares WFB-processed clean audio to WFB-processed noisy audio (worst case scenario)"
[parameters.hearingaid]
name = "SEM Hearing Aid (uFB)"
type = "SEMHearingAid"
processing_strategy = "BatchProcessingOffline"
[parameters.frontend]
name = "uFB"
type = "WFBFrontend"
nbands = 17
fs = 16000.0
spl_reference_db = 100.0
spl_power_estimate_lower_bound_db = 30.0
apcoefficient = 0.0
buffer_size_s = 0.0015
[parameters.backend.general]
name = "SEM"
type = "SEMBackend"
[parameters.backend.inference]
autostart = true
free_energy = false
iterations = 1
[parameters.backend.filters.time_constants90]
s = 5.0 # Speech time constant (ms)
n = 700.0 # Noise time constant (ms)
xnr = 20.0 # ξ time constant (ms)
[parameters.backend.priors.speech]
mean = 80.0
precision = 1.0
[parameters.backend.priors.noise]
mean = 80.0
precision = 1.0
[parameters.backend.gain]
threshold = 12.0 #(GMIN)
[parameters.backend.switch]
threshold = 2.0
[metadata]
author = "VirtualHearingAid"
date = "04-20-2026"
description = "SEM Hearing Aid — uniform filter bank variant (apcoefficient = 0.0) for the WFB ablation reported in the OJ-SP revision."
name = "SEM_uFB"This section presents benchmark results comparing different hearing aid algorithms on the VOICEBANK_DEMAND dataset.
| Device | PESQ (1-5) | DSIG — DNSMOS Signal (1-5) | DBAK — DNSMOS Background (1-5) | DOVRL — DNSMOS Overall (1-5) | CSIG — Composite Signal (1-5) | CBAK — Composite Background (1-5) | COVL — Composite Overall (1-5) |
|---|---|---|---|---|---|---|---|
| SEM | - | - | - | - | - | - | - |
| baseline_clean | - | - | - | - | - | - | - |
| baseline_noise | - | - | - | - | - | - | - |
| SEM_uFB | - | - | - | - | - | - | - |
The following configurations were used for each hearing aid:
[parameters.hearingaid]
name = "SEM Hearing Aid"
type = "SEMHearingAid"
processing_strategy = "BatchProcessingOffline"
[parameters.frontend]
name = "WFB"
type = "WFBFrontend"
nbands = 17
fs = 16000.0
spl_reference_db = 100.0
spl_power_estimate_lower_bound_db = 30.0
apcoefficient = 0.5
buffer_size_s = 0.0015
[parameters.backend.general]
name = "SEM"
type = "SEMBackend"
[parameters.backend.inference]
autostart = true
free_energy = false
iterations = 1
[parameters.backend.filters.time_constants90]
s = 5.0 # Speech time constant (ms)
n = 700.0 # Noise time constant (ms)
xnr = 20.0 # ξ time constant (ms)
[parameters.backend.priors.speech]
mean = 80.0
precision = 1.0
[parameters.backend.priors.noise]
mean = 80.0
precision = 1.0
[parameters.backend.gain]
threshold = 12.0 #(GMIN)
[parameters.backend.switch]
threshold = 2.0
[metadata]
author = "VirtualHearingAid"
date = "03-12-2025"
description = "SEM Hearing Aid configuration"
name = "SEM"
# Baseline Clean Configuration
# This configuration is used for baseline "best" evaluation (clean vs clean)
[metadata]
name = "Baseline Clean"
author = "VirtualHearingAid"
date = "2025-01-27"
description = "Baseline clean evaluation - compares WFB-processed clean audio to itself (best case scenario)"
# Baseline Noise Configuration
# This configuration is used for baseline "worst" evaluation (clean vs noisy)
[metadata]
name = "Baseline Noise"
author = "VirtualHearingAid"
date = "2025-01-27"
description = "Baseline noise evaluation - compares WFB-processed clean audio to WFB-processed noisy audio (worst case scenario)"
[parameters.hearingaid]
name = "SEM Hearing Aid (uFB)"
type = "SEMHearingAid"
processing_strategy = "BatchProcessingOffline"
[parameters.frontend]
name = "uFB"
type = "WFBFrontend"
nbands = 17
fs = 16000.0
spl_reference_db = 100.0
spl_power_estimate_lower_bound_db = 30.0
apcoefficient = 0.0
buffer_size_s = 0.0015
[parameters.backend.general]
name = "SEM"
type = "SEMBackend"
[parameters.backend.inference]
autostart = true
free_energy = false
iterations = 1
[parameters.backend.filters.time_constants90]
s = 5.0 # Speech time constant (ms)
n = 700.0 # Noise time constant (ms)
xnr = 20.0 # ξ time constant (ms)
[parameters.backend.priors.speech]
mean = 80.0
precision = 1.0
[parameters.backend.priors.noise]
mean = 80.0
precision = 1.0
[parameters.backend.gain]
threshold = 12.0 #(GMIN)
[parameters.backend.switch]
threshold = 2.0
[metadata]
author = "VirtualHearingAid"
date = "04-20-2026"
description = "SEM Hearing Aid — uniform filter bank variant (apcoefficient = 0.0) for the WFB ablation reported in the OJ-SP revision."
name = "SEM_uFB"