Rec-RIR: Monaural Blind RIR Identification

• The paper introduces Rec-RIR, a method that estimates monaural RIR by predicting a short STFT-domain convolutive transfer function instead of regressing a long time-domain filter.
• It employs a deep neural network with separate denoising and dereverberation modules to accurately reconstruct the noise-free reverberant speech spectrum.
• The method converts the estimated CTF into a time-domain RIR using a pseudo intrusive measurement process inspired by swept-sine RIR measurement techniques.

Searching arXiv for the specified topic and related papers. Rec-RIR is a method for monaural blind room impulse response identification that estimates a room impulse response from a single-channel noisy reverberant speech recording by first predicting a short-time Fourier transform (STFT)-domain convolutive transfer function (CTF) and then converting that CTF into a time-domain RIR through a pseudo intrusive measurement procedure. Rather than directly regressing a long time-domain filter, it trains a deep neural network through reconstructing the noise-free reverberant speech spectrum, and reports state-of-the-art performance in both RIR identification and acoustic parameter estimation on the reported benchmark (Wang et al., 19 Sep 2025).

1. Problem definition and scope

Rec-RIR addresses the blind monaural setting in which only one observed reverberant channel is available at test time. The observation model is

$y(n) = x(n) + w(n) = h(n) * s(n) + w(n),$

where $y(n)$ is the observed recording, $x(n)$ is the noise-free reverberant speech, $w(n)$ is additive noise, $s(n)$ is clean speech, and $h(n)$ is the room impulse response. The method assumes, for simplicity, that the RIR starts exactly at the sample of the direct-path impulse (Wang et al., 19 Sep 2025).

The central design choice is to avoid direct time-domain regression of $h(n)$. The paper positions this against two prior directions: direct time-domain RIR estimators, which must predict long filters, and STFT-domain methods that estimate clean speech and room filter jointly through iterative optimization or inference. Rec-RIR instead uses a DNN to directly estimate the STFT-domain CTF filter in one forward pass and only afterward converts that estimate to a time-domain RIR. The approach is explicitly described as offline-oriented, and its narrow-band temporal modeling uses bidirectional Mamba because “RIR identification is generally an offline task” (Wang et al., 19 Sep 2025).

2. STFT-domain formulation and CTF approximation

After STFT analysis, Rec-RIR adopts the convolutive transfer function approximation:

$$Y(f,t) = X(f,t) + W(f,t) \approx \sum_{l=0}^{L-1} H_l(f)\, S(f,t-l) + W(f,t),$$

where $f$ is the frequency-bin index, $t$ is the frame index, and $y(n)$0 is the $y(n)$1-th CTF coefficient at frequency bin $y(n)$2. In compact form,

$y(n)$3

with $y(n)$4, $y(n)$5, and $y(n)$6 denoting convolution along the frame axis (Wang et al., 19 Sep 2025).

The paper’s rationale for the CTF representation is that it replaces a very long fullband time-domain convolution by a much shorter per-frequency narrow-band temporal convolution. Rec-RIR therefore treats RIR identification as direct estimation of $y(n)$7, not $y(n)$8. Its input and output are represented by concatenating real and imaginary parts along the channel dimension:

$y(n)$9

$x(n)$0

This formulation is the basis for the architecture, the reconstruction loss, and the subsequent pseudo intrusive conversion back to the time domain (Wang et al., 19 Sep 2025).

3. Network architecture and reconstruction-based learning

The Rec-RIR network comprises four components: an input module, a denoising module, a dereverberation module, and a CTF estimation module. The backbone is derived from SpatialNet and related variants and combines cross-band and narrow-band processing. The input module is a 1-D convolution along the frame axis with kernel size $x(n)$1, and the embedding dimension is $x(n)$2 (Wang et al., 19 Sep 2025).

The denoising and dereverberation modules are separated deliberately. The denoising branch learns embeddings associated with the noise-free reverberant speech, while the dereverberation branch learns embeddings associated with the clean speech. Cross-band blocks capture frequency interactions, whereas narrow-band blocks use the bidirectional Mamba-based narrow-band block from VINP, with one forward Mamba layer and one backward Mamba layer. The denoising module stacks $x(n)$3 cross-band plus narrow-band blocks; the dereverberation module stacks $x(n)$4 such blocks. Each branch has a decoder composed of two linear layers and LeakyReLU to produce $x(n)$5 and $x(n)$6 (Wang et al., 19 Sep 2025).

The CTF estimation module fuses the two embeddings through a learnable weighted sum using scalar parameters $x(n)$7 and $x(n)$8, then uses only narrow-band blocks because the CTF approximation treats different frequency bins as independent subband filters. It stacks $x(n)$9 narrow-band blocks. A weight block with linear layers, LeakyReLU, and softmax computes frequency-dependent frame weights so that an input of arbitrary duration $w(n)$0 can be aggregated into a fixed-length CTF of length $w(n)$1, which the paper states corresponds to an effective RIR length of about 0.96 s (Wang et al., 19 Sep 2025).

Training is organized around reverberant-spectrum reconstruction rather than direct time-domain RIR supervision. The overall loss is

$w(n)$2

with

$w(n)$3

$w(n)$4

$w(n)$5

The paper defines

$w(n)$6

This design makes the estimated CTF physically accountable: it must explain how clean speech becomes reverberant speech under the CTF model. The full training configuration uses $w(n)$7 and $w(n)$8 (Wang et al., 19 Sep 2025).

4. Pseudo intrusive measurement and implementation pipeline

Rec-RIR does not analytically invert the estimated CTF into a waveform-domain RIR. Instead, it introduces a pseudo intrusive measurement process that mimics standard swept-sine RIR measurement. A logarithmic sine sweep $w(n)$9 is chosen together with an inverse filter $s(n)$0 satisfying

$s(n)$1

If $s(n)$2 is the STFT of $s(n)$3, the STFT of a synthetic measurement signal is approximated by

$s(n)$4

Inverse STFT of $s(n)$5 yields $s(n)$6, and the final estimated RIR is recovered by inverse filtering:

$s(n)$7

The sweep used in the experiments spans 62.5 Hz to 8000 Hz, has duration 8.192 s, and uses 256 samples of fade-in and 128 samples of fade-out to reduce spectral leakage (Wang et al., 19 Sep 2025).

The implementation details are comparatively explicit. STFT analysis and synthesis use square-root Hann windows, window length 512 samples, and 50% overlap, giving $s(n)$8 frequency bins. Input normalization divides the waveform by maximum absolute value. Optimization uses AdamW with cosine annealing with restarts, initial learning rate 0.001, batch size 4, 35 training epochs, and 97,092 samples per epoch. Training segments are 4 s long. The resulting model has 3.1M parameters and reported complexity of 62.2 GFlops/s (Wang et al., 19 Sep 2025).

5. Evaluation, ablations, and reported performance

The main evaluation is conducted on SimACE, introduced in VINP. The test mixtures use clean speech from WSJ0 subset si_et_05, measured RIRs from the Single subset of the ACE Challenge, and noises from the REVERB Challenge test set at 20 dB SNR. The reported reference RT60 range is approximately 0.332 s to 1.22 s. Baselines are FiNS, BUDDy including official pre-trained weights, VINP-TCN+SA+S, and VINP-oSpatialNet (Wang et al., 19 Sep 2025).

The reported metrics are RIR-50 ms RMSE for early reflection estimation and acoustic-parameter errors for RT60, DRR, and C50, each with MAE, RMSE, and Pearson correlation coefficient $s(n)$9. Rec-RIR reports the best result on all listed categories: RIR-50 ms RMSE = 0.040; RT60 MAE = 0.069, RMSE = 0.104, $h(n)$0; DRR MAE = 0.684 dB, RMSE = 0.794 dB, $h(n)$1; and C50 MAE = 0.858 dB, RMSE = 1.019 dB, $h(n)$2 (Wang et al., 19 Sep 2025).

The loss ablation isolates the contribution of the auxiliary reverberant and clean-spectrum branches. Compared with the variant $h(n)$3, the full model improves RT60 MAE from 0.077 to 0.069 and DRR MAE from 1.056 dB to 0.684 dB. The paper concludes that both auxiliary losses help, with especially strong benefit for DRR. It also notes that very early impulses occurring before roughly 2 ms are not fully reconstructed; these are attributed to the measurement system’s own frequency response rather than the learned room filter representation (Wang et al., 19 Sep 2025).

6. Position within the RIR literature

Rec-RIR belongs to a specific part of the RIR literature: explicit blind identification of a time-domain room impulse response from monaural reverberant speech. That scope distinguishes it from several adjacent lines of work. RevRIR learns a joint embedding space for reverberant speech and RIRs and applies the speech-side embedding to room shape classification; it is relevant to latent acoustic representation learning but does not decode $h(n)$4 (Bitterman et al., 2024). “Your U-Net Dereverberation Model is Secretly an RIR Encoder” shows that dereverberation U-Nets contain RIR-dependent latent codes and uses explicit RIR embeddings to improve dereverberation, but it stops at the embedding level rather than reconstructing a full RIR waveform (Khanagha et al., 8 Jun 2026).

Other neighboring methods assume different supervision or acquisition regimes. ActiveRIR addresses active audio-visual exploration for building an environment acoustic model from sparse measurements and optimizes where to collect acoustic samples rather than blind speech-based identification (Somayazulu et al., 2024). RIR-Former performs continuous, grid-free reconstruction of missing RIRs at arbitrary coordinates from sparse measured RIRs and microphone positions (Xu et al., 2 Feb 2026). MiNAF predicts high-fidelity RIRs from source/receiver coordinates and mesh-derived explicit local geometry in a neural acoustic field framework (Si et al., 18 Sep 2025). Materialistic RIR addresses material-conditioned realistic RIR generation from scene images and material masks, emphasizing disentanglement of spatial and material effects (Saad et al., 22 Apr 2026). FRA-RIR is not an identifier at all, but a fast stochastic surrogate for ISM-based RIR simulation intended for large-scale augmentation (Luo et al., 2022).

Within that broader landscape, Rec-RIR occupies the niche between direct waveform regression and iterative latent-variable STFT methods. Its defining contribution is the combination of direct CTF prediction, reverberant-spectrum reconstruction training, and pseudo intrusive CTF-to-RIR conversion in a monaural blind setting (Wang et al., 19 Sep 2025).

7. Limitations and implications

The paper identifies several constraints. First, the method is tied to the CTF approximation and the selected STFT configuration. Second, its evaluation centers on simulated mixtures with measured test RIRs, so generalization to more strongly mismatched real recordings is not fully characterized. Third, the architecture is explicitly offline-oriented, which follows from the use of bidirectional Mamba in the narrow-band blocks. Fourth, very early measurement-system artifacts are not perfectly reconstructed (Wang et al., 19 Sep 2025).

These limitations define the method’s present operating point rather than a general barrier to speech-based RIR identification. A plausible implication is that later work may retain the CTF-domain reconstruction principle while changing the temporal backbone, the CTF-to-RIR conversion step, or the supervision regime. Another plausible implication is that the separation of denoising and dereverberation embeddings provides a reusable inductive bias for tasks that require explicit room-filter recovery rather than only dereverberated speech. In its reported form, however, Rec-RIR is best characterized as an offline, monaural, speech-driven RIR identifier whose main innovation lies in estimating a short STFT-domain room filter and then recovering the waveform-domain RIR through a measurement-inspired conversion process (Wang et al., 19 Sep 2025).