Szu-Wei Fu

Hsi-Che Lin, Chien-Yi Wang, Min-Hung Chen et al. · microsoft-research, nvidia

This technical report describes our QuAVF@NTU-NVIDIA submission to the Ego4D Talking to Me (TTM) Challenge 2023. Based on the observation from the TTM task and the provided dataset, we propose to use two separate models to process the input videos and audio. By doing so, we can utilize all the labeled training data, including those without bounding box labels. Furthermore, we leverage the face quality score from a facial landmark prediction model for filtering noisy face input data. The face quality score is also employed in our proposed quality-aware fusion for integrating the results from two branches. With the simple architecture design, our model achieves 67.4% mean average precision (mAP) on the test set, which ranks first on the leaderboard and outperforms the baseline method by a large margin. Code is available at: https://github.com/hsi-che-lin/Ego4D-QuAVF-TTM-CVPR23

Szu-Wei Fu, Rong Chao, Xuesong Yang et al.

Universal Speech Enhancement (USE) aims to restore speech quality under diverse degradation conditions while preserving signal fidelity. Despite recent progress, key challenges in training target selection, the distortion--perception tradeoff, and data curation remain unresolved. In this work, we systematically address these three overlooked problems. First, we revisit the conventional practice of using early-reflected speech as the dereverberation target and show that it can degrade perceptual quality and downstream ASR performance. We instead demonstrate that time-shifted anechoic clean speech provides a superior learning target. Second, guided by the distortion--perception tradeoff theory, we propose a simple two-stage framework that achieves minimal distortion under a given level of perceptual quality. Third, we analyze the trade-off between training data scale and quality for USE, revealing that training on large uncurated corpora imposes a performance ceiling, as models struggle to remove subtle artifacts. Our method achieves state-of-the-art performance on the URGENT 2025 non-blind test set and exhibits strong language-agnostic generalization, making it effective for improving TTS training data. Model weights are available for download at: https://huggingface.co/nvidia/RE-USE.

Ryandhimas E. Zezario, Dyah A. M. G. Wisnu, Szu-Wei Fu et al. · gatech

In this paper, we introduce GatherMOS, a novel framework that leverages large language models (LLM) as meta-evaluators to aggregate diverse signals into quality predictions. GatherMOS integrates lightweight acoustic descriptors with pseudo-labels from DNSMOS and VQScore, enabling the LLM to reason over heterogeneous inputs and infer perceptual mean opinion scores (MOS). We further explore both zero-shot and few-shot in-context learning setups, showing that zero-shot GatherMOS maintains stable performance across diverse conditions, while few-shot guidance yields large gains when support samples match the test conditions. Experiments on the VoiceBank-DEMAND dataset demonstrate that GatherMOS consistently outperforms DNSMOS, VQScore, naive score averaging, and even learning-based models such as CNN-BLSTM and MOS-SSL when trained under limited labeled-data conditions. These results highlight the potential of LLM-based aggregation as a practical strategy for non-intrusive speech quality evaluation.

Ryandhimas E. Zezario, Szu-wei Fu, Fei Chen et al.

Recently, deep learning (DL)-based non-intrusive speech assessment models have attracted great attention. Many studies report that these DL-based models yield satisfactory assessment performance and good flexibility, but their performance in unseen environments remains a challenge. Furthermore, compared to quality scores, fewer studies elaborate deep learning models to estimate intelligibility scores. This study proposes a multi-task speech intelligibility prediction model, called MTI-Net, for simultaneously predicting human and machine intelligibility measures. Specifically, given a speech utterance, MTI-Net is designed to predict human subjective listening test results and word error rate (WER) scores. We also investigate several methods that can improve the prediction performance of MTI-Net. First, we compare different features (including low-level features and embeddings from self-supervised learning (SSL) models) and prediction targets of MTI-Net. Second, we explore the effect of transfer learning and multi-tasking learning on training MTI-Net. Finally, we examine the potential advantages of fine-tuning SSL embeddings. Experimental results demonstrate the effectiveness of using cross-domain features, multi-task learning, and fine-tuning SSL embeddings. Furthermore, it is confirmed that the intelligibility and WER scores predicted by MTI-Net are highly correlated with the ground-truth scores.

Ke-Han Lu, Zhehuai Chen, Szu-Wei Fu et al.

Recent end-to-end speech language models (SLMs) have expanded upon the capabilities of large language models (LLMs) by incorporating pre-trained speech models. However, these SLMs often undergo extensive speech instruction-tuning to bridge the gap between speech and text modalities. This requires significant annotation efforts and risks catastrophic forgetting of the original language capabilities. In this work, we present a simple yet effective automatic process for creating speech-text pair data that carefully injects speech paralinguistic understanding abilities into SLMs while preserving the inherent language capabilities of the text-based LLM. Our model demonstrates general capabilities for speech-related tasks without the need for speech instruction-tuning data, achieving impressive performance on Dynamic-SUPERB and AIR-Bench-Chat benchmarks. Furthermore, our model exhibits the ability to follow complex instructions derived from LLMs, such as specific output formatting and chain-of-thought reasoning. Our approach not only enhances the versatility and effectiveness of SLMs but also reduces reliance on extensive annotated datasets, paving the way for more efficient and capable speech understanding systems.

Quchen Fu, Szu-Wei Fu, Yaran Fan et al.

Meetings are an essential form of communication for all types of organizations, and remote collaboration systems have been much more widely used since the COVID-19 pandemic. One major issue with remote meetings is that it is challenging for remote participants to interrupt and speak. We have recently developed the first speech interruption analysis model, which detects failed speech interruptions, shows very promising performance, and is being deployed in the cloud. To deliver this feature in a more cost-efficient and environment-friendly way, we reduced the model complexity and size to ship the WavLM_SI model in client devices. In this paper, we first describe how we successfully improved the True Positive Rate (TPR) at a 1% False Positive Rate (FPR) from 50.9% to 68.3% for the failed speech interruption detection model by training on a larger dataset and fine-tuning. We then shrank the model size from 222.7 MB to 9.3 MB with an acceptable loss in accuracy and reduced the complexity from 31.2 GMACS (Giga Multiply-Accumulate Operations per Second) to 4.3 GMACS. We also estimated the environmental impact of the complexity reduction, which can be used as a general guideline for large Transformer-based models, and thus make those models more accessible with less computation overhead.

Zhen Wan, Chao-Han Huck Yang, Jinchuan Tian et al.

We introduce a voice-agentic framework that learns one critical omni-understanding skill: knowing when to trust itself versus when to consult external audio perception. Our work is motivated by a crucial yet counterintuitive finding: naively fine-tuning an omni-model on both speech recognition and external sound understanding tasks often degrades performance, as the model can be easily misled by noisy hypotheses. To address this, our framework, Speech-Hands, recasts the problem as an explicit self-reflection decision. This learnable reflection primitive proves effective in preventing the model from being derailed by flawed external candidates. We show that this agentic action mechanism generalizes naturally from speech recognition to complex, multiple-choice audio reasoning. Across the OpenASR leaderboard, Speech-Hands consistently outperforms strong baselines by 12.1% WER on seven benchmarks. The model also achieves 77.37% accuracy and high F1 on audio QA decisions, showing robust generalization and reliability across diverse audio question answering datasets. By unifying perception and decision-making, our work offers a practical path toward more reliable and resilient audio intelligence.

Ke-Han Lu, Szu-Wei Fu, Chao-Han Huck Yang et al.

Large language models (LLMs) have been widely used as knowledge backbones of Large Audio Language Models (LALMs), yet how much auditory knowledge they encode through text-only pre-training and how this affects downstream performance remains unclear. We study this gap by comparing different LLMs under two text-only and one audio-grounded setting: (1) direct probing on AKB-2000, a curated benchmark testing the breadth and depth of auditory knowledge; (2) cascade evaluation, where LLMs reason over text descriptions from an audio captioner; and (3) audio-grounded evaluation, where each LLM is fine-tuned into a Large Audio Language Model (LALM) with an audio encoder. Our findings reveal that auditory knowledge varies substantially across families, and text-only results are strongly correlated with audio performance. Our work provides empirical grounding for a comprehensive understanding of LLMs in audio research.

Pin-Yen Huang, Szu-Wei Fu, Yu Tsao

State-of-the-art (SOTA) semi-supervised learning techniques, such as FixMatch and it's variants, have demonstrated impressive performance in classification tasks. However, these methods are not directly applicable to regression tasks. In this paper, we present RankUp, a simple yet effective approach that adapts existing semi-supervised classification techniques to enhance the performance of regression tasks. RankUp achieves this by converting the original regression task into a ranking problem and training it concurrently with the original regression objective. This auxiliary ranking classifier outputs a classification result, thus enabling integration with existing semi-supervised classification methods. Moreover, we introduce regression distribution alignment (RDA), a complementary technique that further enhances RankUp's performance by refining pseudo-labels through distribution alignment. Despite its simplicity, RankUp, with or without RDA, achieves SOTA results in across a range of regression benchmarks, including computer vision, audio, and natural language processing tasks. Our code and log data are open-sourced at https://github.com/pm25/semi-supervised-regression.

Mirco Ravanelli, Titouan Parcollet, Peter Plantinga et al.

SpeechBrain is an open-source and all-in-one speech toolkit. It is designed to facilitate the research and development of neural speech processing technologies by being simple, flexible, user-friendly, and well-documented. This paper describes the core architecture designed to support several tasks of common interest, allowing users to naturally conceive, compare and share novel speech processing pipelines. SpeechBrain achieves competitive or state-of-the-art performance in a wide range of speech benchmarks. It also provides training recipes, pretrained models, and inference scripts for popular speech datasets, as well as tutorials which allow anyone with basic Python proficiency to familiarize themselves with speech technologies.

Rong Chao, Wen-Huang Cheng, Moreno La Quatra et al. · gatech

This work aims to investigate the use of a recently proposed, attention-free, scalable state-space model (SSM), Mamba, for the speech enhancement (SE) task. In particular, we employ Mamba to deploy different regression-based SE models (SEMamba) with different configurations, namely basic, advanced, causal, and non-causal. Furthermore, loss functions either based on signal-level distances or metric-oriented are considered. Experimental evidence shows that SEMamba attains a competitive PESQ of 3.55 on the VoiceBank-DEMAND dataset with the advanced, non-causal configuration. A new state-of-the-art PESQ of 3.69 is also reported when SEMamba is combined with Perceptual Contrast Stretching (PCS). Compared against Transformed-based equivalent SE solutions, a noticeable FLOPs reduction up to ~12% is observed with the advanced non-causal configurations. Finally, SEMamba can be used as a pre-processing step before automatic speech recognition (ASR), showing competitive performance against recent SE solutions.

Ke-Han Lu, Zhehuai Chen, Szu-Wei Fu et al. · mit

We introduce DeSTA2.5-Audio, a general-purpose Large Audio Language Model (LALM) designed for robust auditory perception and instruction-following, without requiring task-specific audio instruction-tuning. Recent LALMs typically augment Large Language Models (LLMs) with auditory capabilities by training on large-scale, manually curated or LLM-synthesized audio-instruction datasets. However, these approaches have often suffered from the catastrophic forgetting of the LLM's original language abilities. To address this, we revisit the data construction pipeline and propose DeSTA, a self-generated cross-modal alignment strategy in which the backbone LLM generates its own training targets. This approach preserves the LLM's native language proficiency while establishing effective audio-text alignment, thereby enabling zero-shot generalization without task-specific tuning. Using DeSTA, we construct DeSTA-AQA5M, a large-scale, task-agnostic dataset containing 5 million training samples derived from 7,000 hours of audio spanning 50 diverse datasets, including speech, environmental sounds, and music. DeSTA2.5-Audio achieves state-of-the-art or competitive performance across a wide range of audio-language benchmarks, including Dynamic-SUPERB, MMAU, SAKURA, Speech-IFEval, and VoiceBench. Comprehensive comparative studies demonstrate that our self-generated strategy outperforms widely adopted data construction and training strategies in both auditory perception and instruction-following capabilities. Our findings underscore the importance of carefully designed data construction in LALM development and offer practical insights for building robust, general-purpose LALMs.

Szu-Wei Fu, Kuo-Hsuan Hung, Yu Tsao et al.

Speech quality estimation has recently undergone a paradigm shift from human-hearing expert designs to machine-learning models. However, current models rely mainly on supervised learning, which is time-consuming and expensive for label collection. To solve this problem, we propose VQScore, a self-supervised metric for evaluating speech based on the quantization error of a vector-quantized-variational autoencoder (VQ-VAE). The training of VQ-VAE relies on clean speech; hence, large quantization errors can be expected when the speech is distorted. To further improve correlation with real quality scores, domain knowledge of speech processing is incorporated into the model design. We found that the vector quantization mechanism could also be used for self-supervised speech enhancement (SE) model training. To improve the robustness of the encoder for SE, a novel self-distillation mechanism combined with adversarial training is introduced. In summary, the proposed speech quality estimation method and enhancement models require only clean speech for training without any label requirements. Experimental results show that the proposed VQScore and enhancement model are competitive with supervised baselines. The code will be released after publication.

Yen-Ting Lin, Chao-Han Huck Yang, Zhehuai Chen et al.

Construction of a general-purpose post-recognition error corrector poses a crucial question: how can we most effectively train a model on a large mixture of domain datasets? The answer would lie in learning dataset-specific features and digesting their knowledge in a single model. Previous methods achieve this by having separate correction language models, resulting in a significant increase in parameters. In this work, we present Mixture-of-Experts as a solution, highlighting that MoEs are much more than a scalability tool. We propose a Multi-Task Correction MoE, where we train the experts to become an ``expert'' of speech-to-text, language-to-text and vision-to-text datasets by learning to route each dataset's tokens to its mapped expert. Experiments on the Open ASR Leaderboard show that we explore a new state-of-the-art performance by achieving an average relative $5.0$% WER reduction and substantial improvements in BLEU scores for speech and translation tasks. On zero-shot evaluation, NeKo outperforms GPT-3.5 and Claude-Opus with $15.5$% to $27.6$% relative WER reduction in the Hyporadise benchmark. NeKo performs competitively on grammar and post-OCR correction as a multi-task model.

Kuo-Hsuan Hung, Xugang Lu, Szu-Wei Fu et al.

Linguistic knowledge plays a crucial role in spoken language comprehension. It provides essential semantic and syntactic context for speech perception in noisy environments. However, most speech enhancement (SE) methods predominantly rely on acoustic features to learn the mapping relationship between noisy and clean speech, with limited exploration of linguistic integration. While text-informed SE approaches have been investigated, they often require explicit speech-text alignment or externally provided textual data, constraining their practicality in real-world scenarios. Additionally, using text as input poses challenges in aligning linguistic and acoustic representations due to their inherent differences. In this study, we propose the Cross-Modality Knowledge Transfer (CMKT) learning framework, which leverages pre-trained large language models (LLMs) to infuse linguistic knowledge into SE models without requiring text input or LLMs during inference. Furthermore, we introduce a misalignment strategy to improve knowledge transfer. This strategy applies controlled temporal shifts, encouraging the model to learn more robust representations. Experimental evaluations demonstrate that CMKT consistently outperforms baseline models across various SE architectures and LLM embeddings, highlighting its adaptability to different configurations. Additionally, results on Mandarin and English datasets confirm its effectiveness across diverse linguistic conditions, further validating its robustness. Moreover, CMKT remains effective even in scenarios without textual data, underscoring its practicality for real-world applications. By bridging the gap between linguistic and acoustic modalities, CMKT offers a scalable and innovative solution for integrating linguistic knowledge into SE models, leading to substantial improvements in both intelligibility and enhancement performance.

Chih-Kai Yang, Yen-Ting Piao, Tzu-Wen Hsu et al.

Knowledge editing offers an efficient way to update model knowledge without full retraining, but prior work has concentrated almost exclusively on textual or visual modalities. We introduce SAKE, the first benchmark specifically designed for editing auditory attribute knowledge in Large Audio-Language Models (LALMs). Unlike factual updates, SAKE targets several abstract auditory attributes, capturing knowledge types that go beyond conventional textual and visual domains. We benchmark seven editing methods on two LALMs along four dimensions: reliability, generality, audio/text locality, and portability. Results highlight challenges such as preserving intra-attribute knowledge unrelated to the edit, generalizing edits to multimodal reasoning, and maintaining edits under sequential updates. SAKE provides a principled framework to study how knowledge editing extends to the auditory modalities, opening new directions for maintaining and adapting LALMs in more diverse real-world scenarios.

Bo-Han Feng, Chien-Feng Liu, Yu-Hsuan Li Liang et al.

Large audio-language models (LALMs) extend text-based LLMs with auditory understanding, offering new opportunities for multimodal applications. While their perception, reasoning, and task performance have been widely studied, their safety alignment under paralinguistic variation remains underexplored. This work systematically investigates the role of speaker emotion. We construct a dataset of malicious speech instructions expressed across multiple emotions and intensities, and evaluate several state-of-the-art LALMs. Our results reveal substantial safety inconsistencies: different emotions elicit varying levels of unsafe responses, and the effect of intensity is non-monotonic, with medium expressions often posing the greatest risk. These findings highlight an overlooked vulnerability in LALMs and call for alignment strategies explicitly designed to ensure robustness under emotional variation, a prerequisite for trustworthy deployment in real-world settings.

Hung-Chun Hsu, Yuan-Ching Kuo, Chao-Han Huck Yang et al.

The rapid evolution of e-commerce has exposed the limitations of traditional product retrieval systems in managing complex, multi-turn user interactions. Recent advances in multimodal generative retrieval -- particularly those leveraging multimodal large language models (MLLMs) as retrievers -- have shown promise. However, most existing methods are tailored to single-turn scenarios and struggle to model the evolving intent and iterative nature of multi-turn dialogues when applied naively. Concurrently, test-time scaling has emerged as a powerful paradigm for improving large language model (LLM) performance through iterative inference-time refinement. Yet, its effectiveness typically relies on two conditions: (1) a well-defined problem space (e.g., mathematical reasoning), and (2) the model's ability to self-correct -- conditions that are rarely met in conversational product search. In this setting, user queries are often ambiguous and evolving, and MLLMs alone have difficulty grounding responses in a fixed product corpus. Motivated by these challenges, we propose a novel framework that introduces test-time scaling into conversational multimodal product retrieval. Our approach builds on a generative retriever, further augmented with a test-time reranking (TTR) mechanism that improves retrieval accuracy and better aligns results with evolving user intent throughout the dialogue. Experiments across multiple benchmarks show consistent improvements, with average gains of 14.5 points in MRR and 10.6 points in nDCG@1.

Sung-Feng Huang, Heng-Cheng Kuo, Zhehuai Chen et al.

Neural speech editing advancements have raised concerns about their misuse in spoofing attacks. Traditional partially edited speech corpora primarily focus on cut-and-paste edits, which, while maintaining speaker consistency, often introduce detectable discontinuities. Recent methods, like A\textsuperscript{3}T and Voicebox, improve transitions by leveraging contextual information. To foster spoofing detection research, we introduce the Speech INfilling Edit (SINE) dataset, created with Voicebox. We detailed the process of re-implementing Voicebox training and dataset creation. Subjective evaluations confirm that speech edited using this novel technique is more challenging to detect than conventional cut-and-paste methods. Despite human difficulty, experimental results demonstrate that self-supervised-based detectors can achieve remarkable performance in detection, localization, and generalization across different edit methods. The dataset and related models will be made publicly available.

Ke-Han Lu, Zhehuai Chen, Szu-Wei Fu et al.

Recent speech language models (SLMs) typically incorporate pre-trained speech models to extend the capabilities from large language models (LLMs). In this paper, we propose a Descriptive Speech-Text Alignment approach that leverages speech captioning to bridge the gap between speech and text modalities, enabling SLMs to interpret and generate comprehensive natural language descriptions, thereby facilitating the capability to understand both linguistic and non-linguistic features in speech. Enhanced with the proposed approach, our model demonstrates superior performance on the Dynamic-SUPERB benchmark, particularly in generalizing to unseen tasks. Moreover, we discover that the aligned model exhibits a zero-shot instruction-following capability without explicit speech instruction tuning. These findings highlight the potential to reshape instruction-following SLMs by incorporating rich, descriptive speech captions.

Rong Chao, Cheng Yu, Szu-Wei Fu et al.

Speech enhancement (SE) performance has improved considerably owing to the use of deep learning models as a base function. Herein, we propose a perceptual contrast stretching (PCS) approach to further improve SE performance. The PCS is derived based on the critical band importance function and is applied to modify the targets of the SE model. Specifically, the contrast of target features is stretched based on perceptual importance, thereby improving the overall SE performance. Compared with post-processing-based implementations, incorporating PCS into the training phase preserves performance and reduces online computation. Notably, PCS can be combined with different SE model architectures and training criteria. Furthermore, PCS does not affect the causality or convergence of SE model training. Experimental results on the VoiceBank-DEMAND dataset show that the proposed method can achieve state-of-the-art performance on both causal (PESQ score = 3.07) and noncausal (PESQ score = 3.35) SE tasks.

Cheng Yu, Szu-Wei Fu, Tsun-An Hsieh et al.

Although deep learning (DL) has achieved notable progress in speech enhancement (SE), further research is still required for a DL-based SE system to adapt effectively and efficiently to particular speakers. In this study, we propose a novel meta-learning-based speaker-adaptive SE approach (called OSSEM) that aims to achieve SE model adaptation in a one-shot manner. OSSEM consists of a modified transformer SE network and a speaker-specific masking (SSM) network. In practice, the SSM network takes an enrolled speaker embedding extracted using ECAPA-TDNN to adjust the input noisy feature through masking. To evaluate OSSEM, we designed a modified Voice Bank-DEMAND dataset, in which one utterance from the testing set was used for model adaptation, and the remaining utterances were used for testing the performance. Moreover, we set restrictions allowing the enhancement process to be conducted in real time, and thus designed OSSEM to be a causal SE system. Experimental results first show that OSSEM can effectively adapt a pretrained SE model to a particular speaker with only one utterance, thus yielding improved SE results. Meanwhile, OSSEM exhibits a competitive performance compared to state-of-the-art causal SE systems.

Yu-Chen Lin, Cheng Yu, Yi-Te Hsu et al.

Numerous compression and acceleration strategies have achieved outstanding results on classification tasks in various fields, such as computer vision and speech signal processing. Nevertheless, the same strategies have yielded ungratified performance on regression tasks because the nature between these and classification tasks differs. In this paper, a novel sign-exponent-only floating-point network (SEOFP-NET) technique is proposed to compress the model size and accelerate the inference time for speech enhancement, a regression task of speech signal processing. The proposed method compressed the sizes of deep neural network (DNN)-based speech enhancement models by quantizing the fraction bits of single-precision floating-point parameters during training. Before inference implementation, all parameters in the trained SEOFP-NET model are slightly adjusted to accelerate the inference time by replacing the floating-point multiplier with an integer-adder. For generalization, the SEOFP-NET technique is introduced to different speech enhancement tasks in speech signal processing with different model architectures under various corpora. The experimental results indicate that the size of SEOFP-NET models can be significantly compressed by up to 81.249% without noticeably downgrading their speech enhancement performance, and the inference time can be accelerated to 1.212x compared with the baseline models. The results also verify that the proposed SEOFP-NET can cooperate with other efficiency strategies to achieve a synergy effect for model compression. In addition, the just noticeable difference (JND) was applied to the user study experiment to statistically analyze the effect of speech enhancement on listening. The results indicate that the listeners cannot facilely differentiate between the enhanced speech signals processed by the baseline model and the proposed SEOFP-NET.

Ryandhimas E. Zezario, Szu-Wei Fu, Fei Chen et al.

In this study, we propose a cross-domain multi-objective speech assessment model called MOSA-Net, which can estimate multiple speech assessment metrics simultaneously. Experimental results show that MOSA-Net can improve the linear correlation coefficient (LCC) by 0.026 (0.990 vs 0.964 in seen noise environments) and 0.012 (0.969 vs 0.957 in unseen noise environments) in perceptual evaluation of speech quality (PESQ) prediction, compared to Quality-Net, an existing single-task model for PESQ prediction, and improve LCC by 0.021 (0.985 vs 0.964 in seen noise environments) and 0.047 (0.836 vs 0.789 in unseen noise environments) in short-time objective intelligibility (STOI) prediction, compared to STOI-Net (based on CRNN), an existing single-task model for STOI prediction. Moreover, MOSA-Net, originally trained to assess objective scores, can be used as a pre-trained model to be effectively adapted to an assessment model for predicting subjective quality and intelligibility scores with a limited amount of training data. Experimental results show that MOSA-Net can improve LCC by 0.018 (0.805 vs 0.787) in mean opinion score (MOS) prediction, compared to MOS-SSL, a strong single-task model for MOS prediction. In light of the confirmed prediction capability, we further adopt the latent representations of MOSA-Net to guide the speech enhancement (SE) process and derive a quality-intelligibility (QI)-aware SE (QIA-SE) approach accordingly. Experimental results show that QIA-SE provides superior enhancement performance compared with the baseline SE system in terms of objective evaluation metrics and qualitative evaluation test. For example, QIA-SE can improve PESQ by 0.301 (2.953 vs 2.652 in seen noise environments) and 0.18 (2.658 vs 2.478 in unseen noise environments) over a CNN-based baseline SE model.

Szu-Wei Fu, Cheng Yu, Kuo-Hsuan Hung et al.

Most of the deep learning-based speech enhancement models are learned in a supervised manner, which implies that pairs of noisy and clean speech are required during training. Consequently, several noisy speeches recorded in daily life cannot be used to train the model. Although certain unsupervised learning frameworks have also been proposed to solve the pair constraint, they still require clean speech or noise for training. Therefore, in this paper, we propose MetricGAN-U, which stands for MetricGAN-unsupervised, to further release the constraint from conventional unsupervised learning. In MetricGAN-U, only noisy speech is required to train the model by optimizing non-intrusive speech quality metrics. The experimental results verified that MetricGAN-U outperforms baselines in both objective and subjective metrics.

Szu-Wei Fu, Cheng Yu, Tsun-An Hsieh et al.

The discrepancy between the cost function used for training a speech enhancement model and human auditory perception usually makes the quality of enhanced speech unsatisfactory. Objective evaluation metrics which consider human perception can hence serve as a bridge to reduce the gap. Our previously proposed MetricGAN was designed to optimize objective metrics by connecting the metric with a discriminator. Because only the scores of the target evaluation functions are needed during training, the metrics can even be non-differentiable. In this study, we propose a MetricGAN+ in which three training techniques incorporating domain-knowledge of speech processing are proposed. With these techniques, experimental results on the VoiceBank-DEMAND dataset show that MetricGAN+ can increase PESQ score by 0.3 compared to the previous MetricGAN and achieve state-of-the-art results (PESQ score = 3.15).

Ryandhimas E. Zezario, Szu-Wei Fu, Chiou-Shann Fuh et al.

The calculation of most objective speech intelligibility assessment metrics requires clean speech as a reference. Such a requirement may limit the applicability of these metrics in real-world scenarios. To overcome this limitation, we propose a deep learning-based non-intrusive speech intelligibility assessment model, namely STOI-Net. The input and output of STOI-Net are speech spectral features and predicted STOI scores, respectively. The model is formed by the combination of a convolutional neural network and bidirectional long short-term memory (CNN-BLSTM) architecture with a multiplicative attention mechanism. Experimental results show that the STOI score estimated by STOI-Net has a good correlation with the actual STOI score when tested with noisy and enhanced speech utterances. The correlation values are 0.97 and 0.83, respectively, for the seen test condition (the test speakers and noise types are involved in the training set) and the unseen test condition (the test speakers and noise types are not involved in the training set). The results confirm the capability of STOI-Net to accurately predict the STOI scores without referring to clean speech.

Tsun-An Hsieh, Cheng Yu, Szu-Wei Fu et al.

Speech enhancement (SE) aims to improve speech quality and intelligibility, which are both related to a smooth transition in speech segments that may carry linguistic information, e.g. phones and syllables. In this study, we propose a novel phone-fortified perceptual loss (PFPL) that takes phonetic information into account for training SE models. To effectively incorporate the phonetic information, the PFPL is computed based on latent representations of the wav2vec model, a powerful self-supervised encoder that renders rich phonetic information. To more accurately measure the distribution distances of the latent representations, the PFPL adopts the Wasserstein distance as the distance measure. Our experimental results first reveal that the PFPL is more correlated with the perceptual evaluation metrics, as compared to signal-level losses. Moreover, the results showed that the PFPL can enable a deep complex U-Net SE model to achieve highly competitive performance in terms of standardized quality and intelligibility evaluations on the Voice Bank-DEMAND dataset.

Yu-Wen Chen, Kuo-Hsuan Hung, You-Jin Li et al.

This study presents a deep learning-based speech signal-processing mobile application known as CITISEN. The CITISEN provides three functions: speech enhancement (SE), model adaptation (MA), and background noise conversion (BNC), allowing CITISEN to be used as a platform for utilizing and evaluating SE models and flexibly extend the models to address various noise environments and users. For SE, a pretrained SE model downloaded from the cloud server is used to effectively reduce noise components from instant or saved recordings provided by users. For encountering unseen noise or speaker environments, the MA function is applied to promote CITISEN. A few audio samples recording on a noisy environment are uploaded and used to adapt the pretrained SE model on the server. Finally, for BNC, CITISEN first removes the background noises through an SE model and then mixes the processed speech with new background noise. The novel BNC function can evaluate SE performance under specific conditions, cover people's tracks, and provide entertainment. The experimental results confirmed the effectiveness of SE, MA, and BNC functions. Compared with the noisy speech signals, the enhanced speech signals achieved about 6\% and 33\% of improvements, respectively, in terms of short-time objective intelligibility (STOI) and perceptual evaluation of speech quality (PESQ). With MA, the STOI and PESQ could be further improved by approximately 6\% and 11\%, respectively. Finally, the BNC experiment results indicated that the speech signals converted from noisy and silent backgrounds have a close scene identification accuracy and similar embeddings in an acoustic scene classification model. Therefore, the proposed BNC can effectively convert the background noise of a speech signal and be a data augmentation method when clean speech signals are unavailable.

Szu-Wei Fu, Chien-Feng Liao, Tsun-An Hsieh et al.

The Transformer architecture has demonstrated a superior ability compared to recurrent neural networks in many different natural language processing applications. Therefore, our study applies a modified Transformer in a speech enhancement task. Specifically, positional encoding in the Transformer may not be necessary for speech enhancement, and hence, it is replaced by convolutional layers. To further improve the perceptual evaluation of the speech quality (PESQ) scores of enhanced speech, the L_1 pre-trained Transformer is fine-tuned using a MetricGAN framework. The proposed MetricGAN can be treated as a general post-processing module to further boost the objective scores of interest. The experiments were conducted using the data sets provided by the organizer of the Deep Noise Suppression (DNS) challenge. Experimental results demonstrated that the proposed system outperformed the challenge baseline, in both subjective and objective evaluations, with a large margin.

Haoyu Li, Szu-Wei Fu, Yu Tsao et al.

The intelligibility of natural speech is seriously degraded when exposed to adverse noisy environments. In this work, we propose a deep learning-based speech modification method to compensate for the intelligibility loss, with the constraint that the root mean square (RMS) level and duration of the speech signal are maintained before and after modifications. Specifically, we utilize an iMetricGAN approach to optimize the speech intelligibility metrics with generative adversarial networks (GANs). Experimental results show that the proposed iMetricGAN outperforms conventional state-of-the-art algorithms in terms of objective measures, i.e., speech intelligibility in bits (SIIB) and extended short-time objective intelligibility (ESTOI), under a Cafeteria noise condition. In addition, formal listening tests reveal significant intelligibility gains when both noise and reverberation exist.

Cheng Yu, Kuo-Hsuan Hung, Syu-Siang Wang et al.

Previous studies have proven that integrating video signals, as a complementary modality, can facilitate improved performance for speech enhancement (SE). However, video clips usually contain large amounts of data and pose a high cost in terms of computational resources and thus may complicate the SE system. As an alternative source, a bone-conducted speech signal has a moderate data size while manifesting speech-phoneme structures, and thus complements its air-conducted counterpart. In this study, we propose a novel multi-modal SE structure in the time domain that leverages bone- and air-conducted signals. In addition, we examine two ensemble-learning-based strategies, early fusion (EF) and late fusion (LF), to integrate the two types of speech signals, and adopt a deep learning-based fully convolutional network to conduct the enhancement. The experiment results on the Mandarin corpus indicate that this newly presented multi-modal (integrating bone- and air-conducted signals) SE structure significantly outperforms the single-source SE counterparts (with a bone- or air-conducted signal only) in various speech evaluation metrics. In addition, the adoption of an LF strategy other than an EF in this novel SE multi-modal structure achieves better results.

Rung-Yu Tseng, Tao-Wei Wang, Szu-Wei Fu et al.

Speech perception is key to verbal communication. For people with hearing loss, the capability to recognize speech is restricted, particularly in a noisy environment or the situations without visual cues, such as lip-reading unavailable via phone call. This study aimed to understand the improvement of vocoded speech intelligibility in cochlear implant (CI) simulation through two potential methods: Speech Enhancement (SE) and Audiovisual Integration. A fully convolutional neural network (FCN) using an intelligibility-oriented objective function was recently proposed and proven to effectively facilitate the speech intelligibility as an advanced denoising SE approach. Furthermore, audiovisual integration is reported to supply better speech comprehension compared to audio-only information. An experiment was designed to test speech intelligibility using tone-vocoded speech in CI simulation with a group of normal-hearing listeners. Experimental results confirmed the effectiveness of the FCN-based denoising SE and audiovisual integration on vocoded speech. Also, it positively recommended that these two methods could become a blended feature in a CI processor to improve the speech intelligibility for CI users under noisy conditions.

Natalie Yu-Hsien Wang, Hsiao-Lan Sharon Wang, Tao-Wei Wang et al.

The combined electric and acoustic stimulation (EAS) has demonstrated better speech recognition than conventional cochlear implant (CI) and yielded satisfactory performance under quiet conditions. However, when noise signals are involved, both the electric signal and the acoustic signal may be distorted, thereby resulting in poor recognition performance. To suppress noise effects, speech enhancement (SE) is a necessary unit in EAS devices. Recently, a time-domain speech enhancement algorithm based on the fully convolutional neural networks (FCN) with a short-time objective intelligibility (STOI)-based objective function (termed FCN(S) in short) has received increasing attention due to its simple structure and effectiveness of restoring clean speech signals from noisy counterparts. With evidence showing the benefits of FCN(S) for normal speech, this study sets out to assess its ability to improve the intelligibility of EAS simulated speech. Objective evaluations and listening tests were conducted to examine the performance of FCN(S) in improving the speech intelligibility of normal and vocoded speech in noisy environments. The experimental results show that, compared with the traditional minimum-mean square-error SE method and the deep denoising autoencoder SE method, FCN(S) can obtain better gain in the speech intelligibility for normal as well as vocoded speech. This study, being the first to evaluate deep learning SE approaches for EAS, confirms that FCN(S) is an effective SE approach that may potentially be integrated into an EAS processor to benefit users in noisy environments.

Jyun-Yi Wu, Cheng Yu, Szu-Wei Fu et al.

Most recent studies on deep learning based speech enhancement (SE) focused on improving denoising performance. However, successful SE applications require striking a desirable balance between denoising performance and computational cost in real scenarios. In this study, we propose a novel parameter pruning (PP) technique, which removes redundant channels in a neural network. In addition, a parameter quantization (PQ) technique was applied to reduce the size of a neural network by representing weights with fewer cluster centroids. Because the techniques are derived based on different concepts, the PP and PQ can be integrated to provide even more compact SE models. The experimental results show that the PP and PQ techniques produce a compacted SE model with a size of only 10.03% compared to that of the original model, resulting in minor performance losses of 1.43% (from 0.70 to 0.69) for STOI and 3.24% (from 1.85 to 1.79) for PESQ. The promising results suggest that the PP and PQ techniques can be used in a SE system in devices with limited storage and computation resources.

Szu-Wei Fu, Chien-Feng Liao, Yu Tsao et al.

Adversarial loss in a conditional generative adversarial network (GAN) is not designed to directly optimize evaluation metrics of a target task, and thus, may not always guide the generator in a GAN to generate data with improved metric scores. To overcome this issue, we propose a novel MetricGAN approach with an aim to optimize the generator with respect to one or multiple evaluation metrics. Moreover, based on MetricGAN, the metric scores of the generated data can also be arbitrarily specified by users. We tested the proposed MetricGAN on a speech enhancement task, which is particularly suitable to verify the proposed approach because there are multiple metrics measuring different aspects of speech signals. Moreover, these metrics are generally complex and could not be fully optimized by Lp or conventional adversarial losses.

Szu-Wei Fu, Chien-Feng Liao, Yu Tsao

Utilizing a human-perception-related objective function to train a speech enhancement model has become a popular topic recently. The main reason is that the conventional mean squared error (MSE) loss cannot represent auditory perception well. One of the typical hu-man-perception-related metrics, which is the perceptual evaluation of speech quality (PESQ), has been proven to provide a high correlation to the quality scores rated by humans. Owing to its complex and non-differentiable properties, however, the PESQ function may not be used to optimize speech enhancement models directly. In this study, we propose optimizing the enhancement model with an approximated PESQ function, which is differentiable and learned from the training data. The experimental results show that the learned surrogate function can guide the enhancement model to further boost the PESQ score (in-crease of 0.18 points compared to the results trained with MSE loss) and maintain the speech intelligibility.

Chen-Chou Lo, Szu-Wei Fu, Wen-Chin Huang et al.

Existing objective evaluation metrics for voice conversion (VC) are not always correlated with human perception. Therefore, training VC models with such criteria may not effectively improve naturalness and similarity of converted speech. In this paper, we propose deep learning-based assessment models to predict human ratings of converted speech. We adopt the convolutional and recurrent neural network models to build a mean opinion score (MOS) predictor, termed as MOSNet. The proposed models are tested on large-scale listening test results of the Voice Conversion Challenge (VCC) 2018. Experimental results show that the predicted scores of the proposed MOSNet are highly correlated with human MOS ratings at the system level while being fairly correlated with human MOS ratings at the utterance level. Meanwhile, we have modified MOSNet to predict the similarity scores, and the preliminary results show that the predicted scores are also fairly correlated with human ratings. These results confirm that the proposed models could be used as a computational evaluator to measure the MOS of VC systems to reduce the need for expensive human rating.

Yi-Te Hsu, Yu-Chen Lin, Szu-Wei Fu et al.

Numerous studies have investigated the effectiveness of neural network quantization on pattern classification tasks. The present study, for the first time, investigated the performance of speech enhancement (a regression task in speech processing) using a novel exponent-only floating-point quantized neural network (EOFP-QNN). The proposed EOFP-QNN consists of two stages: mantissa-quantization and exponent-quantization. In the mantissa-quantization stage, EOFP-QNN learns how to quantize the mantissa bits of the model parameters while preserving the regression accuracy using the least mantissa precision. In the exponent-quantization stage, the exponent part of the parameters is further quantized without causing any additional performance degradation. We evaluated the proposed EOFP quantization technique on two types of neural networks, namely, bidirectional long short-term memory (BLSTM) and fully convolutional neural network (FCN), on a speech enhancement task. Experimental results showed that the model sizes can be significantly reduced (the model sizes of the quantized BLSTM and FCN models were only 18.75% and 21.89%, respectively, compared to those of the original models) while maintaining satisfactory speech-enhancement performance.

Szu-Wei Fu, Yu Tsao, Hsin-Te Hwang et al.

Nowadays, most of the objective speech quality assessment tools (e.g., perceptual evaluation of speech quality (PESQ)) are based on the comparison of the degraded/processed speech with its clean counterpart. The need of a "golden" reference considerably restricts the practicality of such assessment tools in real-world scenarios since the clean reference usually cannot be accessed. On the other hand, human beings can readily evaluate the speech quality without any reference (e.g., mean opinion score (MOS) tests), implying the existence of an objective and non-intrusive (no clean reference needed) quality assessment mechanism. In this study, we propose a novel end-to-end, non-intrusive speech quality evaluation model, termed Quality-Net, based on bidirectional long short-term memory. The evaluation of utterance-level quality in Quality-Net is based on the frame-level assessment. Frame constraints and sensible initializations of forget gate biases are applied to learn meaningful frame-level quality assessment from the utterance-level quality label. Experimental results show that Quality-Net can yield high correlation to PESQ (0.9 for the noisy speech and 0.84 for the speech processed by speech enhancement). We believe that Quality-Net has potential to be used in a wide variety of applications of speech signal processing.

Szu-Wei Fu, Tao-Wei Wang, Yu Tsao et al.

Speech enhancement model is used to map a noisy speech to a clean speech. In the training stage, an objective function is often adopted to optimize the model parameters. However, in most studies, there is an inconsistency between the model optimization criterion and the evaluation criterion on the enhanced speech. For example, in measuring speech intelligibility, most of the evaluation metric is based on a short-time objective intelligibility (STOI) measure, while the frame based minimum mean square error (MMSE) between estimated and clean speech is widely used in optimizing the model. Due to the inconsistency, there is no guarantee that the trained model can provide optimal performance in applications. In this study, we propose an end-to-end utterance-based speech enhancement framework using fully convolutional neural networks (FCN) to reduce the gap between the model optimization and evaluation criterion. Because of the utterance-based optimization, temporal correlation information of long speech segments, or even at the entire utterance level, can be considered when perception-based objective functions are used for the direct optimization. As an example, we implement the proposed FCN enhancement framework to optimize the STOI measure. Experimental results show that the STOI of test speech is better than conventional MMSE-optimized speech due to the consistency between the training and evaluation target. Moreover, by integrating the STOI in model optimization, the intelligibility of human subjects and automatic speech recognition (ASR) system on the enhanced speech is also substantially improved compared to those generated by the MMSE criterion.

Szu-Wei Fu, Ting-yao Hu, Yu Tsao et al.

This paper aims to address two issues existing in the current speech enhancement methods: 1) the difficulty of phase estimations; 2) a single objective function cannot consider multiple metrics simultaneously. To solve the first problem, we propose a novel convolutional neural network (CNN) model for complex spectrogram enhancement, namely estimating clean real and imaginary (RI) spectrograms from noisy ones. The reconstructed RI spectrograms are directly used to synthesize enhanced speech waveforms. In addition, since log-power spectrogram (LPS) can be represented as a function of RI spectrograms, its reconstruction is also considered as another target. Thus a unified objective function, which combines these two targets (reconstruction of RI spectrograms and LPS), is equivalent to simultaneously optimizing two commonly used objective metrics: segmental signal-to-noise ratio (SSNR) and logspectral distortion (LSD). Therefore, the learning process is called multi-metrics learning (MML). Experimental results confirm the effectiveness of the proposed CNN with RI spectrograms and MML in terms of improved standardized evaluation metrics on a speech enhancement task.

Szu-Wei Fu, Yu Tsao, Xugang Lu et al.

This study proposes a fully convolutional network (FCN) model for raw waveform-based speech enhancement. The proposed system performs speech enhancement in an end-to-end (i.e., waveform-in and waveform-out) manner, which dif-fers from most existing denoising methods that process the magnitude spectrum (e.g., log power spectrum (LPS)) only. Because the fully connected layers, which are involved in deep neural networks (DNN) and convolutional neural networks (CNN), may not accurately characterize the local information of speech signals, particularly with high frequency components, we employed fully convolutional layers to model the waveform. More specifically, FCN consists of only convolutional layers and thus the local temporal structures of speech signals can be efficiently and effectively preserved with relatively few weights. Experimental results show that DNN- and CNN-based models have limited capability to restore high frequency components of waveforms, thus leading to decreased intelligibility of enhanced speech. By contrast, the proposed FCN model can not only effectively recover the waveforms but also outperform the LPS-based DNN baseline in terms of short-time objective intelligibility (STOI) and perceptual evaluation of speech quality (PESQ). In addition, the number of model parameters in FCN is approximately only 0.2% compared with that in both DNN and CNN.