Abstract
There has been a growing effort to develop universal speech enhancement (SE) to handle inputs with various speech distortions and recording conditions. The URGENT Challenge series aims to foster such universal SE by embracing a broad range of distortion types, increasing data diversity, and incorporating extensive evaluation metrics. This work introduces the Interspeech 2025 URGENT Challenge, the second edition of the series, to explore several aspects that have received limited attention so far: language dependency, universality for more distortion types, data scalability, and the effectiveness of using noisy training data. We received 32 submissions, where the best system uses a discriminative model, while most other competitive ones are hybrid methods. Analysis reveals some key findings: (i) some generative or hybrid approaches are preferred in subjective evaluations over the top discriminative model, and (ii) purely generative SE models can exhibit language dependency.
Introduction
Speech Enhancement (SE) has traditionally been a "siloed" field, with researchers often building specialized models for denoising, dereverberation, or bandwidth extension. The URGENT Challenge (Universality, Robustness, and Generalizability of speech EnhancemeNT) aims to break these silos by demanding a single model capable of handling multiple simultaneous distortions. The 2025 edition specifically tackles four under-explored frontiers: language dependency, distortion diversity (including wind noise and packet loss), data scalability (up to 60,000 hours), and the effective use of noisy training data.
Methods and Datasets
The challenge organizers provided a rigorous framework where models processed audio at various sampling rates (8kHz to 48kHz) and handled up to five simultaneous distortions. Participants competed in two tracks: Track 1 (limited to ~2.5k hours) and Track 2 (unlimited data, ~60k hours).
The dataset utilized a massive collection of public corpora including LibriVox, LibriTTS, VCTK, WSJ, EARS, and multilingual sets like MLS-HQ and CommonVoice 19.0. To test generalizability, the blind test set included Japanese data—a language not present in the training sets—and distortions from neural codecs like EnCodec and DAC.
Evaluation was exceptionally thorough, using 14 different metrics categorized into non-intrusive (e.g., DNSMOS), intrusive (e.g., PESQ, SDR), downstream-task-independent (e.g., SpeechBERTScore), and downstream-task-dependent (e.g., Character Accuracy via an ASR model). Subjective Mean Opinion Score (MOS) tests were also conducted.
Key Results and Insights
The challenge results revealed a fascinating trade-off between objective accuracy and human preference. While the #1 ranked system (T1) was a large 102M parameter discriminative sub-band RNN, most other top-tier systems were hybrids combining discriminative and generative losses.
The Generative Trade-off
Purely generative models, such as system T13 (based on latent diffusion), achieved the highest subjective scores (MOS of 3.69) because they produce very "clean-sounding" audio. However, they are prone to "hallucinating" content. Under low-SNR conditions, these models may alter the spoken words to sound more like the languages encountered during training.
Language Sensitivity
Discriminative models proved robust across languages. In contrast, generative models showed a significant performance drop on unseen Japanese data. For instance, T13's Character Accuracy (CAcc) plummeted to 36.8% on Japanese, compared to the baseline's 73.0%. This suggests that generative models may attempt to map unfamiliar phonemes onto known linguistic structures.
Data Scaling Paradox
Surprisingly, no system in Track 2 (60k hours) outperformed the best system in Track 1 (2.5k hours). This suggests that simply increasing data volume is insufficient if the data is noisy or if the model architecture is not optimized to leverage such scale.
Conclusion
The Interspeech 2025 URGENT Challenge highlights the complexity of "universal" speech enhancement. While discriminative models lead in objective metrics like SDR (12.66 dB) and PESQ (2.64), human listeners prefer the output of generative models despite their tendency to hallucinate. Future work will focus on expanding subjective tests to multiple languages and developing new protocols to penalize content changes and speaker inconsistency.
Summary updated