URGENT Speech Enhancement Challenge
- URGENT Speech Enhancement Challenge is a benchmark series evolving from audio-visual enhancement to universal SE, covering tasks like denoising, dereverberation, and bandwidth extension.
- The challenge has progressed since 2019 by incorporating diverse distortions, varying sampling frequencies, multilingual data, and both simulated and real noisy recordings.
- Recent iterations emphasize hybrid models that fuse discriminative and generative methods, using comprehensive objective metrics alongside subjective evaluations to address practical SE issues.
The URGENT Speech Enhancement Challenge denotes a sequence of benchmark efforts that study speech enhancement under increasingly heterogeneous conditions. In the literature, the label first appears in a 2019 audio-visual challenge built around a real noisy binaural corpus and subjective evaluation, and from 2024 onward it designates a broader universal speech enhancement series centered on multiple distortion types, multiple sampling frequencies, multilinguality, curated public training resources, and mixed objective–subjective evaluation (Gogate et al., 2019, Zhang et al., 2024, Saijo et al., 29 May 2025, Li et al., 20 Jan 2026). Across these stages, the challenge has served less as a single fixed task than as an evolving benchmark program for comparing discriminative, generative, and hybrid enhancement systems under common conditions.
1. Historical development and challenge scope
The challenge literature describes a clear progression from realistic audio-visual enhancement toward universal speech enhancement. The 2019 paper presents a “first of its kind” audio-visual speech enhancement challenge in real-noisy settings, while the 2024, 2025, and 2026 papers expand URGENT into a benchmark series for universal, robust, and generalizable enhancement across distortions, sampling rates, languages, and speech styles (Gogate et al., 2019, Zhang et al., 2024, Saijo et al., 29 May 2025, Li et al., 20 Jan 2026).
| Edition | Core scope | Distinctive features |
|---|---|---|
| 2019 AV challenge (Gogate et al., 2019) | Audio-visual speech enhancement | Real noisy ASPIRE corpus; binaural speech; subjective MUSHRA baseline |
| URGENT 2024 (Zhang et al., 2024) | Universal SE | Denoising, dereverberation, bandwidth extension, declipping; multiple sampling frequencies; 12 curated metrics |
| Interspeech 2025 (Saijo et al., 29 May 2025) | Second URGENT edition | Seven distortion types; multilingual training and evaluation; two training-scale tracks |
| ICASSP 2026 (Li et al., 20 Jan 2026) | Two-track URGENT | Track 1 universal SE; Track 2 MOS prediction for enhanced speech |
From 2024 onward, the core task is formulated as
where is the desired speech, is a distortion process, and is the enhanced output (Zhang et al., 2024, Saijo et al., 29 May 2025). In the 2024 overview this broader task initially covers denoising, dereverberation, bandwidth extension, and declipping, with inputs at $8$, $16$, $22.05$, $24$, $32$, $44.1$, and 0 kHz (Zhang et al., 2024). The 2025 edition raises the distortion count to seven by adding codec loss, packet loss, and wind noise, and explicitly studies language dependency and data scalability (Saijo et al., 29 May 2025). The 2026 edition further broadens speech diversity to include different ages, accents, whispered speech, singing voice, and emotional expression, and adds a second track for speech quality assessment (Li et al., 20 Jan 2026).
2. Audio-visual precursor: the 2019 real-noisy benchmark
The 2019 URGENT audio-visual challenge is organized around ASPIRE, a “high-quality AV binaural speech corpus recorded in real noisy settings” (Gogate et al., 2019). The noisy recordings were made in the University of Stirling cafeteria and restaurant during busy lunch times, from 11:30 to 1:30, and matched booth recordings were also collected. ASPIRE contains 5 speakers total, 2 male, 3 female, with age range 23 to 55, mixed English accents, and 10,000 utterances overall: 5,000 in real noisy settings and 5,000 in an acoustically isolated booth (Gogate et al., 2019).
The sentence design follows the AV Grid template. Each utterance is a six-word command sentence with the structure command–colour–preposition–letter–digit–adverb, and each speaker produced 1000 utterances per talker in real noisy settings and 1000 utterances per talker in the booth (Gogate et al., 2019). Video was recorded with an Apple iPad-mini2 at 30 frames per second and 1080p, positioned at 90 cm from the speaker, and high-quality binaural audio was recorded with a Zoom H4n Pro recorder at 44.1 kHz using a binaural microphone worn by a listener at approximately 140 cm (Gogate et al., 2019).
The benchmark is notable because the speech is recorded directly in real environments rather than created by additive mixing. The paper emphasizes nonstationary and multi-source background noise, real reverberation, head movement, face obscuration, and the possibility of the Lombard effect (Gogate et al., 2019). Postprocessing used clap-based audio/video synchronization, Gentle, a forced aligner built on Kaldi, manual boundary checking, and video pixelation for privacy (Gogate et al., 2019).
The benchmark task is audio-visual speech enhancement on real noisy binaural speech, with training on Grid speech mixed with ChiME3 noises and testing on ASPIRE. The baseline systems were SEAGN, Spectrum Subtraction (SS), Log-minimum mean square error (LMMSE), audio-only CochleaNet, and AV CochleaNet (Gogate et al., 2019). Training pooled SNRs from −12 dB to 9 dB in 3 dB steps, with 21000 utterances combined for SNR-independent training (Gogate et al., 2019). Evaluation used a MUSHRA-style listening test with 20 native English speakers, 20 randomly selected utterances, and a 0–100 quality scale (Gogate et al., 2019). The principal finding was that AV CochleaNet performed best and outperformed the audio-only CochleaNet baseline, supporting the challenge premise that visual information improves enhancement under reverberant, multi-source, real-noisy conditions (Gogate et al., 2019).
3. URGENT 2024: universal speech enhancement as a benchmark program
The 2024 URGENT challenge redefines speech enhancement as a broader universal problem rather than a denoising-only task. The overview paper explicitly states that the challenge aims to cover different sub-tasks, data diversity and amount, and evaluation metrics, and it broadens the task definition to denoising, dereverberation, bandwidth extension, and declipping (Zhang et al., 2024). It is restricted to single-channel signals in its first edition, while allowing variable sampling frequencies and multiple distortion types within one system (Zhang et al., 2024).
A central design feature is sample-rate handling. The paper proposes either sampling-frequency-independent (SFI) STFT-based design for models such as BSRNN and TF-GridNet, or an upsample-to-48-kHz strategy for fixed-rate models such as Conv-TasNet (Zhang et al., 2024). In the preliminary challenge experiments, the curated pool contains roughly ~1300 hours of speech and ~250 hours of noise, and the fixed simulated benchmark used ~400 hours for training, ~30 hours for validation, and ~15 hours for test, with SNR range -5 dB to 20 dB and reverberation added with probability 0.5 (Zhang et al., 2024).
Evaluation in the initial overview is deliberately broad. The paper reports 12 curated metrics spanning intrusive, non-intrusive, downstream-task-independent, and downstream-task-dependent categories, including POLQA, PESQ, ESTOI, SDR, MCD, LSD, DNSMOS, NISQA, PhnSim, SpeechBERTScore, SpkSim, and WAcc (Zhang et al., 2024). The preliminary baselines are OM-LSA, VoiceFixer, Conv-TasNet, BSRNN, and TF-GridNet, and the reported trend is that the mapping-based discriminative models, especially BSRNN and TF-GridNet, are much stronger universal baselines than masking-based Conv-TasNet, particularly on bandwidth-related restoration (Zhang et al., 2024). The same experiments also show the classic generative tradeoff: VoiceFixer is strong on DNSMOS and NISQA but weak on intrusive and content-preservation measures, exposing hallucination behavior under the broader metric suite (Zhang et al., 2024).
The post-challenge analysis in “Lessons Learned from the URGENT 2024 Speech Enhancement Challenge” extends these observations from benchmark design to benchmark pathology (Zhang et al., 2 Jun 2025). It reports three phases—validation, non-blind test, and blind test—with 1000 samples in each phase, and a blind test containing 500 simulated samples and 500 real recordings, with 21 participating teams in the final blind test (Zhang et al., 2 Jun 2025). The analysis identifies two systematic data problems. First, declared sampling frequency often mismatched effective bandwidth: ~25% of LibriTTS showed mismatch, while ~100% of DNS5 LibriVox speech and ~100% of CommonVoice 11.0 English showed this issue (Zhang et al., 2 Jun 2025). Second, training labels in nominally clean corpora were often noisy: the paper found approximately 2000 speech labels with negative WADA-SNRs (Zhang et al., 2 Jun 2025). It also finds that the hardest conditions were speech overlap, strong wideband background noise or instantaneous noise, and high reverberation, and that nominal SNR is not a reliable proxy for SE difficulty (Zhang et al., 2 Jun 2025). In the same analysis, the overall combined ranking achieved KRCC between MOS rank and the overall challenge rank = 0.73, while UTMOS and SCOREQ showed the highest correlations with MOS (Zhang et al., 2 Jun 2025).
4. Interspeech 2025: multilinguality, seven distortions, and multi-metric ranking
The Interspeech 2025 URGENT Speech Enhancement Challenge is the second edition of the universal SE series (Saijo et al., 29 May 2025). It is explicitly framed around four underexplored issues: language dependency, universality across more distortion types, data scalability, and how to exploit noisy training data (Saijo et al., 29 May 2025). Compared with URGENT 2024, it expands the distortion set from four to seven: additive noise, reverberation, clipping, bandwidth limitation, codec loss, packet loss, and wind noise, with each sample degraded by up to five distortions simultaneously (Saijo et al., 29 May 2025). The challenge uses two training tracks: about 2.5k hours of speech in Track 1 and about 60k hours of speech in Track 2 (Saijo et al., 29 May 2025).
Multilinguality becomes a first-class component of validity. Training covers five languages—English, German, French, Spanish, and Chinese—and the blind test set adds Japanese as an unseen language, with 150 samples per language and 900 total samples, half simulated and half real (Saijo et al., 29 May 2025). The challenge uses 14 metrics grouped into five categories: non-intrusive SE metrics (DNSMOS, NISQA, UTMOS), intrusive SE metrics (POLQA, PESQ, ESTOI, SDR, MCD, LSD), downstream-task-independent metrics (LPS, SBS), downstream-task-dependent metrics (SpkSim, CAcc), and subjective MOS via ITU-T P.808 (Saijo et al., 29 May 2025). The final ranking is not based on a single metric but on a three-step procedure inspired by the Friedman test: rank each metric, average ranks within each category, then average category-wise rankings, with a lower final score indicating a better system (Saijo et al., 29 May 2025).
The 2025 challenge received 32 submissions total. The printed subset of the Track 1 leaderboard shows T1 as the best overall system with final ranking score 2.97, and it is explicitly characterized as purely discriminative; most of the remaining competitive systems are hybrid (Saijo et al., 29 May 2025). The most revealing entry is the purely generative T13. It ranked first on DNSMOS, NISQA, UTMOS, and MOS, with MOS 3.69, yet performed very poorly on intrusive and content-sensitive measures, including POLQA 1.99, PESQ 1.34, ESTOI 0.54, SDR -12.28, SpkSim 0.47, and CAcc 67.87% (Saijo et al., 29 May 2025). The same paper reports strong language dependence for this generative system: Chinese CAcc 20.1 and Japanese CAcc 36.8, despite high DNSMOS, whereas the top discriminative model T1 remains relatively stable across languages (Saijo et al., 29 May 2025). These results establish one of the defining challenge debates: subjective naturalness and non-intrusive quality can diverge sharply from content preservation and speaker fidelity.
5. System families and architectural patterns shaped by URGENT
A distinct research ecosystem has formed around URGENT-style benchmarks, and several recurring design patterns are visible in challenge papers. One line favors regression-first universal models. USEMamba, a state-space model with sampling-frequency-independent feature extraction and a distortion-aware combination of regression and generation, achieved 2nd place in Track 1 during the blind test phase of URGENT 2025 despite being trained on only a subset of the data (Chao et al., 27 May 2025). Its central claim is that regression works well for most distortions, while a generative variant is more suitable for packet loss and bandwidth extension (Chao et al., 27 May 2025).
A second line emphasizes multi-stage fusion. FUSE combines a Sparse Compression Network, a token-based generative refinement stage, and a final fusion network, augmented by a shift trick and output blending; the paper states that this system achieved third rank in the URGENT 2025 challenge (Goswami et al., 1 Jun 2025). In its blind-test comparison among the top three systems, FUSE reports the highest DNSMOS, NISQA, UTMOS, and MOS, while remaining weaker than the best discriminative model on several intrusive metrics (Goswami et al., 1 Jun 2025). A more deployment-oriented variant is DeepFilterGAN, a full-band real-time system with 3.45M inference parameters, 20 ms window, 10 ms hop, 2-frame look-ahead, and 40 ms algorithmic latency, explicitly described as a 2025 URGENT challenge participant (Serbest et al., 29 May 2025).
The 2026 papers reinforce the same architectural convergence toward generative–predictive fusion. The 2026 challenge overview states that the strongest Track 1 systems mostly used hybrid generative + discriminative paradigms, often dual-branch or multi-stage, and that successful teams emphasized MOS-based filtering of training data and quality-aware curation (Li et al., 20 Jan 2026). Representative systems include GAP-URGENet, which combines a generative branch built from DeWavLM-Omni, Adapter, and Vocoder with a predictive TF-GridNet branch and a PostNet for learned fusion and bandwidth extension. That system reports 567.76M parameters, 472.84 GMACs per second, and validation metrics of DNSMOS 3.31, NISQA 3.96, UTMOS 3.22, SCOREQ 4.03, PESQ 2.89, ESTOI 0.90, SBS 0.91, LPS 0.89, and SpkSim 0.82, and it achieved 1st place in the objective evaluation on the ICASSP 2026 blind test (Rong et al., 2 Apr 2026). Another 2026 submission, “A Hybrid Discriminative and Generative System for Universal Speech Enhancement,” combines sampling-frequency-independent TF-GridNet, an autoregressive generative branch with spectral mapping, and TF-domain fusion, and it ranked third place in Track 1 (Liu et al., 27 Jan 2026). In parallel, UniPASE extends the low-hallucination PASE framework with DeWavLM-Omni, an acoustic Adapter, a neural Vocoder, and a PostNet; the full UniPASE stack reports 545.7M parameters and 79.2 GMAC/s, and the paper states that a simple hybrid extension built from UniPASE and predictive TF-GridNet achieved 1st place in the objective evaluation of URGENT 2026 (Rong et al., 16 Apr 2026).
6. Evaluation controversies, data curation, and recurrent research lessons
One of the most influential later papers is the multilingual evaluation study “P.808 Multilingual Speech Enhancement Testing: Approach and Results of URGENT 2025 Challenge” (Sach et al., 15 Jul 2025). It formalizes the crowdsourced ITU-T P.808 ACR pipeline for multilingual MOS testing and localizes it to English, German, Chinese, and Japanese. The challenge MOS uses eight ratings per utterance, and the paper reports first-stage acceptance rates of 82.3% for English, 62.8% for German, 41.1% for Chinese, and 28.9% for Japanese (Sach et al., 15 Jul 2025). Its most important result is that the language ordering by MOS matches the ordering by the phone-fidelity metric
1
but not the ordering by ESTOI or the reference-free metrics (Sach et al., 15 Jul 2025). The paper highlights Japanese as the clearest mismatch case: MOS 2.94, LPS 0.53, yet DNSMOS 2.88 and NISQA 3.08 (Sach et al., 15 Jul 2025). At the model level, the generative model #13 reached MOS 3.34 ± 0.05, DNSMOS 3.13, and NISQA 3.79, while collapsing on ESTOI 0.53 and LPS 0.58 ± 0.03; the authors describe the resulting outputs as exhibiting a “high degree of hallucination” (Sach et al., 15 Jul 2025). The paper’s conclusion is not that MOS is obsolete, but that reference-free subjective ACR MOS and common reference-free objective metrics may jointly fail to detect hallucinations, so they should be accompanied by phone-fidelity metrics such as LPS and, where possible, native listeners (Sach et al., 15 Jul 2025).
A second major line of critique concerns training targets and data quality. The 2026 paper “Rethinking Training Targets, Architectures and Data Quality for Universal Speech Enhancement” argues that the conventional dereverberation target of early-reflected speech is suboptimal and that time-shifted anechoic clean speech is a better target (Fu et al., 3 Mar 2026). On the URGENT 2025 non-blind test set, its preferred configuration, shifted anechoic + GAN, reports DNSMOS 3.26, NISQA 4.12, UTMOS 2.80, and CAcc 89.88 (Fu et al., 3 Mar 2026). The same paper also argues that training on large uncurated corpora imposes a performance ceiling, and it uses VQScore filtering to show that moderate quality filtering gives a better tradeoff than either no filtering or aggressive pruning (Fu et al., 3 Mar 2026). This concern aligns with the 2024 lessons paper on bandwidth mismatch and noisy labels (Zhang et al., 2 Jun 2025) and with the 2025 overview’s observation that simply scaling to ~60k hours in Track 2 did not produce a system that surpassed the best Track 1 submission (Saijo et al., 29 May 2025).
Taken together, these studies show that URGENT is not merely a leaderboard of enhancement models. It has become a testbed for unresolved questions about what speech enhancement should optimize, how challenge data should be curated, and how systems should be ranked when perceptual quality, intelligibility, linguistic fidelity, and speaker preservation are not interchangeable (Sach et al., 15 Jul 2025, Fu et al., 3 Mar 2026, Zhang et al., 2 Jun 2025). A plausible implication is that the challenge’s long-term significance lies as much in its evaluation methodology and dataset governance as in any single winning architecture.
