We investigate how the realism of synthetic room impulse response (RIR) datasets affects the training of DeepFilterNet3 for single-channel speech enhancement. We compare a DNS4 image-source-method (ISM) RIR dataset with a higher-acoustic-fidelity dataset generated using hybrid wave-based and geometrical acoustics simulation. Rather than isolating individual simulation factors, we compare complete RIR generation pipelines while keeping the enhancement model unchanged. Models are evaluated on unseen measured RIRs using objective speech enhancement metrics and downstream automatic speech recognition (ASR). Training with the higher-fidelity dataset consistently yields modest improvements in objective metrics and substantially lower ASR word error rates than the ISM dataset. Although the experiments do not attribute these gains to individual modelling components, they show that increasing the overall realism of synthetic acoustic training data improves the generalization of DeepFilterNet3 to unseen measured environments.
Figures & tables
Fig. 1 : Histogram of Sabine RT60 values for Hybrid and ISM matched datasets
Fig. 2 : Level of detail for one of the rooms simulated in the Hybrid dataset
Parameter
Value
Network version
0.5.7pre
Network type
deepfilternet3
DRR
0.3
RT60target
0.05s
Late reverberation offset
5ms
preverb
1.00
Table 1: DeepFilterNet3 training configuration
Speech
Config
RIR type
Case
Small 30 ep.
Default
Hybrid
1
ISM
2
Small † 30 ep.
No decay aug.
Hybrid
3
ISM
4
Full 30 ep.
Default
Hybrid
5
ISM
6
Table 2: Overview of experimental configurations. Speech samples extracted from DNS4 English dataset with the addition of PTDB-TUG. † RT decay randomization from DFN3 disabled
Setting
Metric
ISM
Hybrid
Δ
Small 30 ep.
PESQ
2.17
2.35
+0.18
SISDRi
1.20
1.41
+0.21
STOI
0.700
0.714
+0.014
SRMR
9.80
9.97
+0.18
Small † 30 ep.
PESQ
2.18
2.39
+0.20
SISDRi
1.30
1.35
+0.05
Table 3: Objective evaluation metrics (median) for DFN3 trained with ISM and Hybrid RIR datasets across different training regimes. † RT decay randomization from DFN3 disabled.
Metric
Δ (Hybrid–ISM)
95% CI
PESQ
+0.166
[+0.149, +0.184]
SISDRi
+0.110
[+0.040, +0.182]
STOI
+0.013
[+0.010, +0.015]
SRMR
+0.270
[+0.213, +0.328]
Table 4: Overall paired improvement of Hybrid over ISM across all configurations. Values are mean differences with 95% bootstrap confidence intervals paired at utterance level.
Model
WER ↓
Δ vs noisy
Rel. imp. (%)
Clean reference
0.0224
+0.1447
–
Hybrid (Treble) training
Full 30 ep.
0.1274
+0.0397
23.8%
Full 120 ep.
0.1290
+0.0381
22.8%
Small † 30 ep.
0.1399
+0.0272
16.3%
Small 30 ep.
0.1470
+0.0201
12.0%
Table 5: ASR performance (WER) using NeMo on enhanced speech. Δ vs noisy denotes the absolute WER reduction with respect to the noisy baseline (0.1671), while Rel. imp. (%) denotes the corresponding relative improvement. Training with the Hybrid (Treble) dataset consistently yields lower WER than training with the ISM (DNS4) dataset across all evaluated configurations. † RT decay randomization from DFN3 disabled.
Speech enhancement improves intelligibility and listening quality by suppressing noise, reverberation, and competing speakers in recorded or live audio.
Room-acoustic simulations are widely used to augment training data for deep-learning-based speech enhancement. While most pipelines rely on simplified geometrical acoustics, wave-based approaches offer greater physical accuracy. In this work, we examine how simulation fidelity affects multichannel speech enhancement performance. To this end, we train SpatialNet on datasets augmented with different room-acoustic simulation methods and evaluate the resulting models on measured data. We compare lower-fidelity datasets based on geometrical acoustics with a high-fidelity dataset using advanced acoustic modelling and a hybrid combination of wave-based and geometrical acoustics simulations. Training on the high-fidelity dataset results in an up to 38 % relative reduction in median word error rate compared to the lower-fidelity alternatives. These results show that augmentation with high-fidelity room-acoustic simulations directly translates into improved multichannel speech enhancement performance.
We propose a self-supervised approach for learning room impulse response (RIR) representations from single-channel noisy-reverberant speech. It consists of first training on reverberant data, then on noisy-reverberant data, and finally with a teacher-student approach, where the student learns to replicate the teacher's embeddings when given a noisy version of the reverberant input. We assess their representational capabilities by estimating acoustic room parameters from them. Conditioning a discriminative speech enhancement model on the derived embeddings yields consistent gains across all evaluated metrics, including downstream word error rate, for both reverberant and noisy-reverberant speech.
We introduce Whisper-RIR-Mega, a benchmark dataset of paired clean and reverberant speech for evaluating automatic speech recognition (ASR) robustness to room acoustics. Each sample pairs a clean LibriSpeech utterance with the same utterance convolved with a real room impulse response from the RIR-Mega corpus, with stratified splits by reverberation time (RT60) and direct-to-reverberant ratio (DRR). We evaluate five Whisper models (tiny through large-v3) on 1600 test samples and report word error rate (WER) and character error rate (CER) under clean and reverberant conditions. Reverberation consistently degrades performance across all model sizes; the reverb penalty in WER ranges from 2.31 to 15.50 percentage points depending on the model. Whisper-large-v3 shows the smallest penalty; Whisper-tiny shows the largest. We release the dataset, evaluation code, and baseline results to support reproducible research on robust ASR.