FAST-RIR: Neural Impulse Response Generator

Updated 5 July 2026

The paper introduces FAST-RIR as a neural network-based RIR generator using a 10-D conditioning vector to simulate both specular and diffuse reflections with high precision.
It leverages a modified Conditional GAN architecture with tailored loss functions, including adversarial, MSE, and T60 consistency terms, to ensure accurate reverberation time reproduction.
Benchmarks indicate FAST-RIR achieves up to 400x speedup over traditional simulators and improves ASR accuracy by significantly reducing word error rates.

FastRIR, introduced as “FAST-RIR: Fast neural diffuse room impulse response generator,” is a neural-network-based fast diffuse room impulse response generator for generating room impulse responses for a given acoustic environment. In its original formulation, it takes rectangular room dimensions, listener and speaker positions, and reverberation time as inputs and generates specular and diffuse reflections for that acoustic environment; a later summary specifies a 10-D embedding comprising rectangular room dimensions $(L, W, H)$ , source position $SP(x,y,z)$ , listener position $LP(x,y,z)$ , and reverberation time $T_{60}$ , with a monaural raw-waveform output of 4096 samples at 16 kHz (Ratnarajah et al., 2021, Ratnarajah, 2024).

1. Origin and problem setting

FAST-RIR was published in October 2021 as a method intended to accelerate room impulse response generation while preserving utility in automatic speech recognition applications (Ratnarajah et al., 2021). The method was evaluated using Google Speech API, Microsoft Speech API, and Kaldi tools, and the original report emphasized both acoustic control and downstream ASR performance.

The problem it addresses is the computational cost of large-scale RIR simulation. In the later technical summary, room impulse responses are described as the transfer function from a source to a receiver in an environment, encoding direct sound, early reflections, and late reverberation; convolving a dry signal with an RIR produces reverberant audio used in ASR, speech enhancement, separation, and related tasks (Ratnarajah, 2024). Against that background, FAST-RIR is positioned as a learning-based alternative to computationally expensive simulators such as a diffuse acoustic simulator (DAS) and as a faster generator than gpuRIR in the reported benchmarks (Ratnarajah et al., 2021).

A central feature of the original method is explicit conditioning on room and source–listener attributes. This distinguishes FAST-RIR from purely stochastic reverberation models: it is not an unconditioned sampler, but a conditional generator intended to produce an RIR matching a specified rectangular-room configuration and target reverberation time.

2. Conditioning, representation, and network formulation

In the configuration summarized for ASR augmentation, FAST-RIR uses a 10-D conditioning vector comprising room dimensions, source position, listener position, and $T_{60}$ (Ratnarajah, 2024). The generator is conditioned on this embedding only, with no noise variable $z$ , so that it produces a single precise RIR for the condition.

The same source describes FAST-RIR as a modified Conditional GAN. Its networks were adapted from Stage-I StackGAN to 1D audio, with enlarged receptive fields, kernel length 41, and stride up to 4 to handle low frequencies; transposed convolutions replace upsampling-plus-convolution (Ratnarajah, 2024). The output is a monaural RIR waveform of 4096 samples at 16 kHz, corresponding to approximately 0.256 s.

The generator objective combines adversarial, waveform, and reverberation-time consistency terms:

$L_{OGAN} = E_{JA\sim P_{data}}[\log(1 - D_N(G_N(JA)))]$

$L_{MSE} = E_{JA\sim P_{data}}[E[(R_N(JA, s) - R_D(JA, s))^2]]$

$L_{T60} = E_{JA\sim P_{data}}[|T60(G_N(JA)) - T60(JA)|]$

and

$L_{GN} = L_{OGAN} + \alpha_{MSE}L_{MSE} + \alpha_{T60}L_{T60}.$

The discriminator objective is reported as

$SP(x,y,z)$ 0

These loss terms show that FAST-RIR was designed not only to imitate target waveforms but also to control the generated reverberation time directly (Ratnarajah, 2024).

The original abstract states that the generator produces both specular and diffuse reflections (Ratnarajah et al., 2021). In this sense, FAST-RIR is not limited to a late-tail approximation; it is presented as a full RIR generator for the specified acoustic environment.

3. Training data and parameter regime

The later summary reports that FAST-RIR was trained on 75,000 DAS-simulated RIRs. The training set covered room dimensions $SP(x,y,z)$ 1 m, $SP(x,y,z)$ 2 m, $SP(x,y,z)$ 3 m, random source and listener positions, and $SP(x,y,z)$ 4 in $SP(x,y,z)$ 5 s (Ratnarajah, 2024).

Training used RMSprop with learning rate $SP(x,y,z)$ 6, batch size 128, and decay 0.7 every 40 epochs (Ratnarajah, 2024). These details indicate that FAST-RIR was developed as a supervised conditional generator trained against simulator-produced targets rather than as a simulator-free acoustics model.

The original paper reports that FAST-RIR is capable of generating RIRs for a given input reverberation time with an average error of 0.02 s (Ratnarajah et al., 2021). A later analysis refines this point for short reverberation times: when input $SP(x,y,z)$ 7 s, cropping the generated RIR at $SP(x,y,z)$ 8 reduces average error from 0.029 s to 0.023 s (Ratnarajah, 2024). This places reverberation-time fidelity at the center of the method’s design and evaluation.

Because the training corpus was DAS-simulated and the conditioning variables explicitly encode a rectangular room, FAST-RIR is best understood as a learned surrogate for a constrained family of room-acoustic simulations rather than as a general-purpose geometry-free acoustics estimator.

4. Runtime characteristics and ASR evaluation

The original abstract reports three headline results: with batch size 1, FAST-RIR is 400 times faster than a state-of-the-art diffuse acoustic simulator on a CPU; it is 12 times faster than gpuRIR; and it outperforms gpuRIR by 2.5% in an AMI far-field ASR benchmark (Ratnarajah et al., 2021). A later summary gives the underlying timing data for 30,000 RIRs (Ratnarajah, 2024).

Method	Total time	Average time per RIR
DAS (CPU)	$SP(x,y,z)$ 9 s	30.05 s
Image Method (CPU)	$LP(x,y,z)$ 0 s	0.15 s
FAST-RIR (CPU, batch=1)	$LP(x,y,z)$ 1 s	0.07 s
gpuRIR (GPU)	16.63 s	$LP(x,y,z)$ 2 s
FAST-RIR (GPU, batch=64)	1.33 s	$LP(x,y,z)$ 3 s

These timings make clear that FAST-RIR was evaluated in both CPU and GPU regimes. The CPU comparison is especially important for augmentation pipelines that do not assume GPU-resident simulation, while the GPU comparison shows that the learned generator can exceed a specialized accelerated simulator in raw throughput.

The ASR results were reported in two forms. For Google and Microsoft API tests on LibriSpeech test-clean convolved with simulated RIRs, the relative WER increase versus a DAS baseline was reported as gpuRIR $LP(x,y,z)$ 4, Image Method $LP(x,y,z)$ 5, and FAST-RIR $LP(x,y,z)$ 6 (Ratnarajah, 2024). On the AMI far-field ASR task using the Kaldi IHM $LP(x,y,z)$ 7SDM recipe, the reported scores were gpuRIR: dev 52.2, eval 55.5; DAS: dev 47.9, eval 52.5; FAST-RIR: dev 47.8, eval 53.0 (Ratnarajah, 2024). The abstract’s statement that FAST-RIR outperforms gpuRIR by 2.5% on the AMI benchmark is consistent with these reported values (Ratnarajah et al., 2021).

Taken together, these results place FAST-RIR in a specific operating regime: it was developed for high-throughput RIR synthesis where the relevant success criterion is not only acoustic plausibility but also the ability of the generated RIRs to support downstream far-field ASR.

5. Later adaptations and downstream reuse

A later use of FastRIR appears in the ICASSP 2025 Room Acoustics and Speaker Distance Estimation Challenge, where the open-source fast diffuse room impulse response generator was modified so that it was conditioned only on speaker and listener locations, not on explicit room geometry or materials (Ratnarajah et al., 1 May 2026). In that work, the generator was extended to output 1-second, 32 kHz, single-channel time-domain RIRs, and its representation was adapted following MESH2IR to preserve consistent energy distribution across different source–receiver distances.

The adaptation used pre-training on 100k GWA RIRs and subsequent fine-tuning on limited Treble and GWA enrollment subsets, with separate generators maintained for Treble and GWA because the two datasets use different simulation methodologies and exhibit different acoustics (Ratnarajah et al., 1 May 2026). The reported generation pipeline sampled many source–receiver pairs, synthesized approximately 1,000,000 RIRs, and then applied a quality filter. The filter accepted only RIRs with reverberation time within $LP(x,y,z)$ 8 of the reference distribution for the target set, rejected any with $LP(x,y,z)$ 9 s, rejected distances below 0.8 m or above 7.1 m, enforced DRR consistency with distance, required a typical approximately exponential energy decay curve, and rejected atypical early reflection patterns (Ratnarajah et al., 1 May 2026).

After filtering, the acceptance yield was approximately 25%, resulting in approximately 260,000 high-quality RIRs (Ratnarajah et al., 1 May 2026). Those accepted RIRs were convolved with dry speech to fine-tune a speaker distance estimation model, reducing mean absolute error from 1.66 m to 0.6 m for GWA rooms and from 2.18 m to 0.69 m for Treble rooms (Ratnarajah et al., 1 May 2026). Dataset-specific SDE models yielded further MAE reductions of 10% for Treble and 5% for GWA relative to a unified model (Ratnarajah et al., 1 May 2026).

This later usage is methodologically significant because it shows that FastRIR could be reconfigured from explicit room-attribute conditioning to position-only conditioning. It also shows the cost of that simplification: reverberation characteristics such as RT60, EDT, DRR, and early/late structure were not explicitly set by the user, but instead controlled indirectly by dataset selection, sampled distance, and post-generation filtering (Ratnarajah et al., 1 May 2026).

6. Relations to adjacent methods, naming, and scope

FastRIR has been used as a reference point in later acoustics work. In “Few-Shot Audio-Visual Learning of Environment Acoustics,” the authors state explicitly that Fast-RIR is a different method from FS-RIR and use an extended “Fast-RIR++” baseline for comparison (Majumder et al., 2022). In that comparison, the original Fast-RIR is characterized as assuming known room attributes such as size and RT60 and rectangular rooms; the extended baseline estimates RT60 and DRR and infers scene size from panoramic depth (Majumder et al., 2022). This is a direct indication of how the original method was perceived in subsequent literature: strong in attribute-conditioned rectangular-room synthesis, but not natively formulated for arbitrary scene geometry or few-shot scene adaptation.

The literature also contains similarly named but distinct methods. FRAM-RIR and FRA-RIR are fast random approximations of the image-source method rather than conditional neural generators (Luo et al., 2023, Luo et al., 2022). FS-RIR is a transformer-based few-shot audio-visual method for predicting arbitrary binaural RIRs in novel environments from sparse egocentric images and echoes (Majumder et al., 2022). An unrelated paper in wireless communications uses “FastRIR” as the name of a framework for fast reconfiguration of LC-RISs (Delbari et al., 11 Apr 2025). These naming collisions have practical importance in bibliographic search.

The scope of the original FAST-RIR is correspondingly precise. It is a neural-network-based fast diffuse RIR generator for rectangular rooms, explicitly conditioned on room dimensions, source and listener positions, and reverberation time, and evaluated primarily through ASR augmentation performance (Ratnarajah et al., 2021). A plausible implication is that its strongest regime is large-scale generation when the target acoustics can be described by those conditioning variables. By contrast, later works that removed explicit room conditioning relied on dataset priors and filtering, while later few-shot and mesh-based methods targeted broader scene variability through different representations (Ratnarajah et al., 1 May 2026, Majumder et al., 2022).