Achieving high perceptual quality without hallucination remains a challenge in generative speech enhancement (SE). A representative approach, PASE, is robust to hallucination but has limited perceptual quality under adverse conditions. We propose StuPASE, built upon PASE to achieve studio-level quality while retaining its low-hallucination property. First, we show that finetuning PASE with dry targets rather than targets containing simulated early reflections substantially improves dereverberation. Second, to address performance limitations under strong additive noise, we replace the GAN-based generative module in PASE with a flow-matching module, enabling studio-quality generation even under highly challenging conditions. Experiments demonstrate that StuPASE consistently produces perceptually high-quality speech while maintaining low hallucination, outperforming state-of-the-art SE methods. Audio demos are available at: https://xiaobin-rong.github.io/stupase_demo/.
Figures & tables
Figure 1: Overview of the proposed StuPASE framework.
Model
UTMOS ↑
SpkSim ↑
dWER (%) ↓
Noisy
1.30
0.70
10.23
Clean
4.14
1.00
0.00
DeWavLM
2.42
0.54
10.30
DeWavLM-R
3.98
0.55
8.69
PASE
1.61
0.74
9.78
PASE-R
3.23
0.80
8.01
Table 1: Ablation on DNS1 with-reverb test set. Blue-shaded rows are for references; † denotes inference on clean speech.
Model
No-Reverb
With-Reverb
DNSMOS ↑
UTMOS ↑
SBS ↑
LPS ↑
SpkSim ↑
dWER (%) ↓
DNSMOS ↑
UTMOS ↑
SBS ↑
LPS ↑
SpkSim ↑
dWER (%) ↓
Noisy
2.48
2.36
0.80
0.90
0.96
3.51
1.39
1.30
0.61
0.63
0.79
10.23
Clean
3.28
4.14
1.00
1.00
1.00
0.00
3.28
4.14
1.00
1.00
1.00
0.00
TF-GridNet [ 38 ]
3.34
3.86
0.91
0.97
0.94
2.86
2.63
1.42
0.77
0.88
0.80
8.86
FlowSE [ 2 ]
3.38
3.76
0.90
0.94
0.89
4.65
3.34
3.51
0.81
0.85
0.72
15.58
PASE [ 13 ]
3.29
3.95
0.93
0.97
0.88 ∗
2.71
2.75
1.61
0.81
0.90
0.74 ∗
9.78
Table 2: Comparison results on the DNS1 test set. Bold indicates the best scores, and underline indicates the second-best.
Model
DNSMOS ↑
UTMOS ↑
SBS ↑
LPS ↑
SpkSim ↑
WER (%) ↓
Noisy
1.34
1.42
0.56
0.53
0.54
14.74
Clean
3.35
4.28
1.00
1.00
1.00
3.14
TF-GridNet
2.87
2.22
0.77
0.82
0.73
13.46
FlowSE
3.28
3.50
0.76
0.74
0.59
27.84
PASE
2.96
2.44
0.82
0.86
0.53
13.95
SenSE
3.39
3.88
0.83
0.88
0.68
12.73
Table 3: Comparison results on the simulated test set.
Figure 2: Subjective evaluation results for Q-MOS and S-MOS, with error bars indicating 95% confidence intervals.