RawTFNet: Cross-Domain Raw Signal Model

• RawTFNet is a design philosophy that employs raw signal processing with explicit time–frequency separation to capture modality-specific features.
• In speech anti-spoofing, a lightweight CNN leverages axis-specific convolutions on raw waveforms to efficiently detect subtle synthetic artifacts.
• In radar applications, learnable Fourier transforms and transformer backbones offer a data-adaptive alternative to traditional FFT-based pipelines.

Searching arXiv for RawTFNet and related papers to ground the article in the cited literature. RawTFNet is a term used in more than one recent arXiv context, and its meaning is therefore domain-dependent rather than canonical. In speech anti-spoofing, "RawTFNet" denotes a lightweight convolutional neural network that operates end-to-end on raw audio and separates feature processing along time and frequency dimensions to detect spoofing artifacts (Xiao et al., 11 Jul 2025). In automotive radar, the term does not appear as the paper’s formal model name, but T-FFTRadNet is explicitly described as instantiating a “Raw Transformer Network” for FMCW radar because it processes raw ADC signals through learnable complex transforms and a hierarchical Swin Transformer backbone (Giroux et al., 2023). By contrast, TFNet in neural speech coding refers to a time–frequency speech codec rather than a raw-signal RawTFNet formulation (Jiang et al., 2022).

1. Terminological scope and usage

The term "RawTFNet" is not associated with a single universally standardized architecture. In the speech anti-spoofing literature, it is the proper name of a model introduced for automatic speaker verification countermeasures, with a design centered on raw-waveform input and explicit time–frequency factorization inside a lightweight CNN (Xiao et al., 11 Jul 2025). In the radar literature, the closest corresponding usage is conceptual rather than titular: T-FFTRadNet is presented as a transformer-based detector operating on raw ADC radar data through learnable Fourier-like layers, and is described as a Raw Transformer Network for radar (Giroux et al., 2023).

This suggests that "RawTFNet" functions as a family resemblance term rather than a single architecture: it denotes systems that begin from raw sensor data and impose a structured time–frequency or Fourier-domain inductive bias within the network. A plausible implication is that the shared label reflects methodological convergence across modalities—raw-input modeling, learned front ends, and reduced dependence on fixed preprocessing—rather than architectural identity.

A related source of confusion is the earlier speech-coding model TFNet. That model is an end-to-end neural speech codec for real-time communications, but it does not operate directly in the raw time domain; instead, it processes STFT magnitude and phase using an encoder–temporal filtering–decoder pipeline (Jiang et al., 2022). The same source explicitly notes that the term "RawTFNet" does not appear there.

2. RawTFNet in speech anti-spoofing

In speech anti-spoofing, RawTFNet is a lightweight CNN architecture designed for audio signals and evaluated on ASVspoof 2021 LA and DF (Xiao et al., 11 Jul 2025). Its stated motivation is the need to detect subtle synthetic-speech artifacts while avoiding the compute demands of recent transformer-based systems such as Rawformer and AASIST. The model addresses this by explicitly separating feature processing along time and frequency dimensions, with the claim that such separation is well matched to the axis-specific statistical signatures of spoofing artifacts (Xiao et al., 11 Jul 2025).

The system is end-to-end on raw waveforms sampled at 16 kHz. Input segments are 4 s, corresponding to 64,000 samples. A learnable Sinc convolutional layer with 70 trainable band-pass filters converts the 1D waveform into a 2D time–frequency feature map, eliminating the need for hand-crafted STFT or filterbank front ends (Xiao et al., 11 Jul 2025). This learned frontend is followed by four lightweight residual blocks implemented as depthwise-separable SE-Res2Blocks. The first block uses 32 channels, while the next three use 64 channels; these blocks incorporate Res2Net-style multi-scale sub-branches with dilation and a squeeze-and-excitation module, with batch normalization and ReLU in all layers (Xiao et al., 11 Jul 2025).

The architectural core is a stack of nine identical TF-Conv blocks. Each block first applies a transition $1 \times 1$ pointwise convolution and channel shuffle, then splits channels into frequency and time branches of equal size. The frequency branch applies frequency-only depthwise convolution with kernel $3 \times 1$, frequency-wise average pooling, and a $1 \times 1$ projection; the time branch analogously applies time-only depthwise convolution with kernel $1 \times 3$, time-wise pooling, and pointwise projection (Xiao et al., 11 Jul 2025). The two enhanced outputs are concatenated, with BN and ReLU used in all sublayers. Two $2 \times 2$ max-pooling layers with stride 2 are inserted at intermediate depths, and classification is performed by global average pooling over frequency and time followed by a $1 \times 1$ convolution that produces a binary bonafide-versus-spoof logit (Xiao et al., 11 Jul 2025).

The paper gives the branch-wise broadcast-add operations as

$\hat{x}_f = \sum_{j=1}^{T} (x_{ij,f} + v_{j,f}),$

and

$\hat{x}_t = \sum_{i=1}^{F} (x_{ij,t} + v_{i,t}),$

with $v_f$ and $v_t$ denoting 1D summaries derived from pooling and pointwise projection (Xiao et al., 11 Jul 2025). Frequency-only and time-only depthwise convolutions are likewise formalized as axis-specific 1D operators over a tensor $3 \times 1$0, followed by $3 \times 1$1 pointwise projection (Xiao et al., 11 Jul 2025).

Training uses the ASVspoof 2019 LA training set, Adam with learning rate $3 \times 1$2, weight decay $3 \times 1$3, batch size 32, and 100 epochs. Checkpoint averaging over the top 5 validation checkpoints is used for evaluation. The loss is weighted cross-entropy,

$3 \times 1$4

and RawBoost “series: (1+2+3)” (Algo4) is applied as augmentation (Xiao et al., 11 Jul 2025).

3. Performance and ablations in anti-spoofing

Two width-scaled variants are reported. RawTFNet-16 uses 0.07M parameters and 2.9G MACs, while RawTFNet-32 uses 0.17M parameters and 5.4G MACs (Xiao et al., 11 Jul 2025). On ASVspoof 2021 LA, RawTFNet-16 achieves 4.50% EER and 0.295 min t-DCF, whereas RawTFNet-32 reaches 5.05% EER and 0.321 min t-DCF. On ASVspoof 2021 DF, RawTFNet-16 yields 18.53% EER and RawTFNet-32 yields 16.82% EER (Xiao et al., 11 Jul 2025).

The paper positions these results against several baselines. Reported comparisons include RawGT-ST at 0.44M parameters and 17.4G MACs with LA EER 10.25% and DF EER 23.26%; RawNet2 at 17.60M parameters and 0.8G MACs with LA EER 9.50% and DF EER 22.38%; AASIST at 0.30M parameters and 8.9G MACs with LA EER 10.51%, min t-DCF 0.488, and DF EER 21.07%; SE-Rawformer at 0.37M parameters and 6.1G MACs with LA EER 4.98%, min t-DCF 0.318, and DF EER 20.26%; and Rawformer-L at 0.29M parameters and 9.2G MACs with LA EER 6.83% and min t-DCF 0.361 (Xiao et al., 11 Jul 2025). On that basis, the paper states that RawTFNet reaches comparable performance to state-of-the-art systems while using fewer computing resources.

Ablation results on ASVspoof 2021 LA, reported for RawTFNet-32, identify the TF-Conv decomposition as the dominant design factor. The full model obtains EER 5.05% with 5400M MACs and 170K parameters. Removing the frequency branch increases EER to 6.91% with 4550M MACs and 150K parameters, while removing the temporal branch increases EER to 7.08% with 4480M MACs and 150K parameters. Removing channel shuffle yields EER 5.74% with unchanged 5400M MACs and 170K parameters (Xiao et al., 11 Jul 2025). These figures support the paper’s claim that both axis-specific branches are crucial and that channel shuffle contributes materially to feature interaction.

The study also reports duration sensitivity. EER is worst for utterances of 0–2 s, improves in the 2–4 s range, rises slightly for 4–8 s, and then drops again for durations above 8 s, albeit with fewer samples (Xiao et al., 11 Jul 2025). This suggests that training on 4 s segments influences the operating regime of the detector.

4. RawTFNet as a radar Raw Transformer Network

In automotive radar, the most direct counterpart to RawTFNet is T-FFTRadNet, introduced for object detection from FMCW radar signals and explicitly characterized as a “Raw Transformer Network” because it replaces standard FFT preprocessing with learnable complex transformations and uses a hierarchical Swin Transformer backbone (Giroux et al., 2023). The problem setting is radar object detection from raw ADC signals acquired across fast-time samples, slow-time chirps, and multiple Tx/Rx antennas. Conventional pipelines compute range, Doppler, and angle FFTs to produce Range–Azimuth–Doppler cubes; the paper identifies multiple FFTs, windowing, normalization, and CPU-side preprocessing as computational burdens, and further notes that fixed transforms may discard sensor characteristics that are valuable for learning, especially with low-definition radars (Giroux et al., 2023).

T-FFTRadNet operates directly on raw complex ADC inputs through learnable complex linear transformations that mimic FFTs. Each transformation layer is an $3 \times 1$5 complex linear operator initialized as a DFT,

$3 \times 1$6

and trained end-to-end so that it can deviate from the pure Fourier basis and learn data- and sensor-specific projections (Giroux et al., 2023). An fftshift-like reindexing is applied to center zero Doppler, and permutations between layers orient multiplication along range or Doppler axes (Giroux et al., 2023). The paper also evaluates an optional phase-preserving complex nonlinearity,

$3 \times 1$7

reported as “ADC–NL” (Giroux et al., 2023).

The transformed outputs are normalized and then tokenized for a hierarchical Swin Vision Transformer. Radar inputs are treated as pseudo-images whose spatial axes correspond to range and Doppler for ADC/RD inputs or range and angle for RAD inputs, with antennas and I/Q components placed in the channel dimension (Giroux et al., 2023). The architecture uses $3 \times 1$8 patch embedding, windowed multi-head self-attention with shifted windows, and patch merging between stages. No explicit positional encoding is added; the paper states that Swin’s windowing and shifted-window mechanism provide implicit relative positional structure (Giroux et al., 2023).

Downstream of the Swin backbone, a Range–Angle Decoder maps compressed RD features into RA space, following FFTRadNet’s strategy of decoding angular information without an explicit azimuth FFT (Giroux et al., 2023). The detector uses an anchor-free binary RA grid to predict object centers at reduced resolution, with a regression head for fine range and angle correction. For low-definition radar, a classification head predicts per-pixel class likelihood maps using Softmax and pixel-wise cross-entropy; for high-definition radar, a free-driving-space segmentation head uses binary BCE on a polar RA grid (Giroux et al., 2023).

5. Radar signal model, losses, and datasets

The radar formulation is grounded in standard FMCW relations. The paper gives the range resolution and maximum unambiguous range as

$3 \times 1$9

and

$1 \times 1$0

and the Doppler shift as

$1 \times 1$1

It also specifies conventional range, Doppler, and angle FFTs and notes that stacking them yields an RAD cube with axes $1 \times 1$2 (Giroux et al., 2023). These formulas are not novel to T-FFTRadNet, but they establish the signal-processing background against which the learned transform layers are positioned.

For input organization, the HD RadIal dataset uses 12 Tx and 16 Rx, with 256 chirps per transmitter and 512 samples per chirp. Inputs are ADC or RD tensors of shape $1 \times 1$3, where complex I/Q are appended as additional channels, or RAD cubes using magnitudes only (Giroux et al., 2023). The LD RADDet dataset uses 2 Tx, 4 Rx, 64 chirps, and 256 samples, with inputs of shape $1 \times 1$4 after virtual array formation; again, ADC/RD inputs treat complex components as channels and RAD cubes use magnitudes (Giroux et al., 2023).

Conventional RD/RAD pipelines use per-channel normalization,

$1 \times 1$5

and may apply Hamming windows over range and Doppler axes (Giroux et al., 2023). By contrast, raw ADC pipelines inject unnormalized inputs into the learned transforms and normalize only after those transforms and before the Swin backbone (Giroux et al., 2023).

The detection and auxiliary tasks are trained through a linear combination of losses. Binary detection plus localization uses

$1 \times 1$6

with reduction factors $1 \times 1$7, where HD uses $1 \times 1$8, LD uses $1 \times 1$9, and $1 \times 3$0, $1 \times 3$1 (Giroux et al., 2023). LD classification uses

$1 \times 3$2

with $1 \times 3$3, and HD free-space segmentation uses

$1 \times 3$4

with $1 \times 3$5 (Giroux et al., 2023).

Training is reported for PyTorch 1.12.1 on a single NVIDIA RTX 3090 with an Intel i7-12700K, using 150 epochs, Adam, learning rate $1 \times 3$6, and decay 0.9 every 10 epochs. Batch sizes are 4 for HD, 16 for LD(RD), and 4 for LD(RAD) (Giroux et al., 2023).

6. Comparative performance and practical implications

On HD RadIal, T-FFTRadNet achieves approximately 89.5% F1, 89.5% AR, and 89.6% AP, with free-space mIoU approximately 80.2%, range error about 0.15 m, and angle error about $1 \times 3$7 (Giroux et al., 2023). The free-space result is reported as on par with Cross-Modal DNN at 80.4% and significantly above FFTRadNet at 73.98% (Giroux et al., 2023). Raw ADC variants perform slightly lower, with ADC at about 87.4% F1 and ADC–NL at about 87.1% F1; windowing improves metrics, and Doppler fftshift yields marginal changes in HD (Giroux et al., 2023).

On LD RADDet, the model attains 55.7% mAP on RAD cubes versus 51.6% for RADDet, 46.8% mAP on RD maps comparable to DAROD at 46.6%, and 49.1% mAP on raw ADC, with ADC–NL at 45.6% (Giroux et al., 2023). For generalized detection, the reported AP/AR values are 64.5%/57.5% for RD, 64.2%/56.5% for RAD, and 61.8%/54.5% for ADC (Giroux et al., 2023). Class-wise AP/AR shows underperformance on motorcycles, which the paper attributes to dataset imbalance and signature similarity (Giroux et al., 2023).

A central practical result concerns preprocessing cost and latency. On RadIal, T-FFTRadNet uses 194 G FLOPs and 9.64 M parameters, compared with FFTRadNet’s 288 G FLOPs and 3.79 M parameters and Cross-Modal DNN’s 358 G FLOPs and 7.7 M parameters (Giroux et al., 2023). For Fourier preprocessing overhead, the RadIal RD FFT is reported at 214 M FLOPs, whereas the learned Fourier-Net block for ADC transformation costs 12.9 G FLOPs with 327.68 K parameters; on RADDet, RD FFT costs 11 M FLOPs and Fourier-Net 337.6 M FLOPs with 69.63 K parameters (Giroux et al., 2023). Despite this learned-transform cost, end-to-end inference time favors the raw-ADC path in the reported setup: on HD, RD plus normalization is approximately 156 ms versus raw ADC approximately 30 ms; on LD, RAD is approximately 152 ms, RD approximately 33 ms, and ADC approximately 20 ms (Giroux et al., 2023). The paper attributes this to learned transform layers running efficiently on GPU while conventional FFTs and normalization were CPU-bound in the tests.

These results indicate that the RawTFNet principle can trade fixed signal-processing stages for trainable front ends without necessarily increasing end-to-end latency. In radar, the benefit appears strongest in low-definition settings, where learned sensor-adaptive projections can compensate for reduced resolution (Giroux et al., 2023). In speech anti-spoofing, the benefit is realized as a sharp efficiency–accuracy trade-off achieved by axis-specific convolutions over raw audio (Xiao et al., 11 Jul 2025).

7. Relationship to TFNet and broader methodological significance

TFNet for real-time neural speech coding belongs to a related but distinct methodological lineage. It uses a time–frequency front end based on STFT with a 20 ms window and 5 ms hop, causal 2D convolutions in the encoder and decoder, interleaved temporal filtering modules composed of dilated temporal convolution modules and group-wise GRUs, and a VQ stage with bitrate

$1 \times 3$8

for example $1 \times 3$9, $2 \times 2$0, $2 \times 2$1, $2 \times 2$2 at 6 kbps (Jiang et al., 2022). The model is optimized jointly for coding, speech enhancement, and packet loss concealment, and operates with approximately 20 ms algorithmic delay (Jiang et al., 2022). However, the paper explicitly distinguishes TFNet from any raw-waveform RawTFNet interpretation, stating that TFNet processes STFT magnitude and phase rather than raw time-domain signals (Jiang et al., 2022).

The comparison clarifies two separate uses of the initials "TF". In TFNet, "TF" denotes time–frequency processing in a codec (Jiang et al., 2022). In RawTFNet for anti-spoofing, the name refers to raw-audio modeling with time–frequency-separated convolutions (Xiao et al., 11 Jul 2025). In radar, the RawTFNet label is best understood as an interpretive description of T-FFTRadNet’s raw-input transformer design, not as the paper’s official architecture name (Giroux et al., 2023).

Across these usages, a common pattern emerges: the move away from rigid handcrafted front ends toward learned transforms that retain explicit structural priors. In speech anti-spoofing, those priors are axis-separated time and frequency convolutions over a raw-audio-derived representation (Xiao et al., 11 Jul 2025). In radar, they are complex DFT-initialized layers aligned with range and Doppler structure, followed by transformer tokenization (Giroux et al., 2023). This suggests that RawTFNet is best regarded not as a single model family with a fixed blueprint, but as a cross-domain design philosophy in which raw-signal ingestion, structured transform learning, and computational efficiency are jointly emphasized.