Research map / Architectures and efficiency / Attention and transformers

TIGER: Time-frequency Interleaved Gain Extraction and Reconstruction for Efficient Speech Separation

ICLR 2025 · Architectures and efficiency › Attention and transformers · approach: Long-context KV and memory attention · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TIGER, a lightweight time-frequency speech separation model using band-split, multi-scale selective attention and full-frequency-frame attention, plus the EchoSet dataset with realistic reverberation from SoundSpaces 2.0 and Matterport3D.

Problem

Speech separation models are either accurate but computationally heavy (e.g., TF-GridNet) or efficient but weaker, and common benchmarks (WSJ0-2mix, Libri2Mix, LRS2-2Mix) do not capture realistic reverberation, limiting real-world generalization.

Approach

Band-split the frequency axis using prior knowledge (narrow low-frequency bands, wide high-frequency bands) to compress frequency information, then interleave a multi-scale selective attention module (U-Net-like with selective fusion) and a full-frequency-frame attention module across frequency and frame paths with shared parameters.

Key results

• TIGER (large) reaches 14.22 dB SDRi on EchoSet, above TF-GridNet's 13.73 dB, while using 94.3% fewer parameters (0.82M vs 14.43M) and 95.3% fewer MACs.
• On Libri2Mix and LRS2-2Mix, TIGER (large) trails TF-GridNet by roughly 1.2 and 0.4 dB SDRi (18.34 vs 19.56; 15.3 vs 15.7).
• On cinematic sound separation (DnR), TIGER reaches 15.5 dB on speech with 1.40M parameters and 4.07 G MACs/s.

Measured effects reported in the paper

Surprising findings

• Models trained on EchoSet generalized better to real-world recordings than those trained on Libri2Mix and LRS2-2Mix.
• Replacing the frame path with a frequency path only (F-F) drops SDRi sharply to 10.57 dB, versus 13.15 dB for the interleaved F-T design.

Limitations

• On Libri2Mix and LRS2-2Mix, TIGER underperforms TF-GridNet in separation quality.
• Real-world evaluation set is small-scale (10 rooms, about 1.5 hours of speech), and EchoSet covers only two-speaker mixtures.
• Dataset construction relies on simulated room acoustics for training, with the real-world test only validating generalization.

Experimental setup

Datasets: EchoSet Libri2Mix LRS2-2Mix LibriSpeech WHAM! noise DnR

Benchmarks: EchoSet Libri2Mix LRS2-2Mix DnR real-world recordings

Metrics: SDRi SI-SDRi parameters MACs GPU time CPU time GPU memory

Baselines: Conv-TasNet DualPathRNN SudoRM-RF1.0x A-FRCNN-16 TDANet Large BSRNN TF-GridNet SepFormer

Builds on: TF-GridNet TDANet BSRNN DPRNN SudoRM-RF

Closest papers by meaning

• RTFS-Net: Recurrent Time-Frequency Modelling for Efficient Audio-Visual Speech Separation (ICLR 2024 · similarity 0.87)
• Separate and Reconstruct: Asymmetric Encoder-Decoder for Speech Separation (NeurIPS 2024 · similarity 0.86)
• WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling (ICLR 2025 · similarity 0.84)
• Weakly-supervised Audio Separation via Bi-modal Semantic Similarity (ICLR 2024 · similarity 0.83)
• NaturalSpeech 2: Latent Diffusion Models are Natural and Zero-Shot Speech and Singing Synthesizers (ICLR 2024 · similarity 0.83)
• E-BATS: Efficient Backpropagation-Free Test-Time Adaptation for Speech Foundation Models (NeurIPS 2025 · similarity 0.83)