DPCRN: Dual-Path Convolution Recurrent Network for Single Channel Speech Enhancement

1. Introduction

• Low SNR, high reverberation, and far-field pickup remain challenges for DNN-based enhancement in real applications.The DNS challenge provides training datasets and evaluation metrics to foster systems for adverse environments.Source: §1 · p. 1
• DPRNN addresses the difficulty of modeling extremely long sequences by splitting them into smaller chunks processed with intra-chunk and inter-chunk RNNs.This reduces the sequence length handled by each RNN.Source: §1 · p. 1
• DPCRN combines CRN and DPRNN in the time-frequency domain, using intra-chunk RNNs for single-frame spectra and inter-chunk RNNs for temporal spectral variation.The model outputs a complex ratio mask from its decoder.Source: §1 · pp. 1–2
• With 0.8M parameters, DPCRN achieves an overall MOS of 3.57 and ranks third in the DNS challenge wide-band track.On the DNS challenge blind test set, the subjective evaluation followed ITU-T P.835.Source: §1 · p. 2

2.1. Problem formulation

• The noisy time-frequency spectrogram is enhanced by multiplying it element-wise by an estimated mask, and signal approximation minimizes the difference from clean speech.The paper uses a complex ratio mask and converts the enhanced spectrogram back to waveform with inverse STFT.Source: §2.1 · p. 2, §2.3 · pp. 2–3

2.2. Model architecture

• The encoder-decoder extracts local spectrogram patterns, reduces and restores feature resolution, and uses skip connections to pass detailed information.The DPRNN replaces the CRN recurrent component: intra-chunk RNNs model spectral patterns within a frame, while causal inter-chunk LSTMs model temporal dependence at each frequency.Source: §2.2 · p. 2
• Instant layer normalization calculates statistics independently for each frame over frequency and channel axes, while sharing trainable parameters.The model also applies iLN to the input spectrogram to reduce output sensitivity to input signal energy.Source: §2.2 · p. 2

2.3. Learning Target and Loss function

• The training target is the complex ratio mask, whose real and imaginary parts are decoder outputs, and training optimizes signal approximation.The loss combines negative SNR with spectrogram MSE; the MSE measures real, imaginary, and magnitude differences and is logarithmically scaled to match the negative-SNR loss magnitude.Source: §2.3 · pp. 2–3

3.1. Datasets

• Training uses 60,000 reverberant-speech clips, with 55,000 for training and 5,000 for validation; simulated tests use WSJ0 speech with MUSAN or NOISEX92 noise.Training mixtures use randomly selected room impulse responses and SNRs from -5 to 5 dB; test mixtures use the same SNR range.Source: §3.1 · p. 3

3.2. Parameter setup

• The causal model uses two DPRNN modules, about 0.8M parameters, and 37.5 ms latency; DPCRN-1, -2, and -3 vary loss choice or DPRNN input frequency resolution.DPCRN-3 halves that frequency resolution while doubling the intra-chunk RNN hidden dimension for the same computational complexity.Source: §3.2 · p. 3

3.3. Baselines and evaluation metrics

• Comparisons include DTLN, DCCRN, and NSNet2, using PESQ, STOI, and SDR on simulated WSJ0, DNSMOS on the development set, and subjective P.835 evaluation on the blind test.The blind-test evaluation is the final result.Source: §3.3 · pp. 3–4

3.4. Results and analysis

• DPCRN-1 is slightly weaker than DCCRN at SNRs of at least 0 dB but better at lower SNR, and exceeds baselines on all three metrics under NOISEX92 noise.On WSJ0-MUSAN, DPCRN-1 is also better than DTLN at SNRs of at least 0 dB.Source: Table 1 · p. 4, Table 2 · p. 4, §3.4 · p. 4
• On the DNS development set, DPCRN models with about 0.8M parameters and no look-ahead outperform baselines on DNSMOS; the submitted model earns overall MOS 3.57 and third place.The submitted model has a little speech quality deterioration; the development-set DPCRN-2 is the best of the compared models.Source: Table 3 · p. 4, Table 4 · p. 4, §3.4 · p. 4
• DPCRN-2 improves PESQ and STOI over DPCRN-1 on both simulated datasets but has slightly worse SDR.The authors attribute the speech-quality benefit to including time-frequency MSE in the loss.Source: §3.4 · p. 4
• DPCRN-3 has more parameters than DPCRN-1 yet performs worse; reducing DPRNN input frequency resolution is detrimental, while the submitted model uses dimension 50 for resolution and real-time requirements.RNN parallel computation is a challenge for real-time processing.Source: §3.4 · p. 4

• With only 0.8M parameters, DPCRN achieves competitive results on various unknown-noise datasets.Source: §4 · p. 4
• DPCRN combines CNN local-pattern modeling with DPRNN long-term modeling for time-frequency speech enhancement.Source: §4 · p. 4
• Compared with CRN, DPCRN demonstrates the benefit of RNN for spectrum modeling.Source: §4 · p. 4
• The authors identify reducing computational complexity as future work for wider-band spectrum processing.Source: §4 · p. 4