Search

Language diversity presents a significant challenge in speech-to-text (S2T) tasks, such as automatic speech recognition and translation. Traditional multi-task training approaches aim to address this by jointly optimizing multiple speech recognition and translation tasks across various languages. While models like Whisper, built on these strategies, demonstrate strong performance, they still face issues of high computational cost, language interference, suboptimal training configurations, and limited extensibility. To overcome these challenges, we introduce LoRS-Merging (low-rank and sparse model merging), a novel technique designed to efficiently integrate models trained on different languages or tasks while preserving performance and reducing computational overhead. LoRS-Merging combines low-rank and sparse pruning to retain essential structures while eliminating redundant parameters, mitigating language and task interference, and enhancing extensibility. Experimental results across a range of languages demonstrate that LoRS-Merging reduces the word error rate by 10% and improves BLEU scores by 4% compared to conventional multi-lingual multi-task training baselines. Our findings suggest that model merging, particularly LoRS-Merging, is a scalable and effective complement to traditional multi-lingual training strategies for S2T applications., Comment: 13 pages, submitted to ACL 2025

We present a novel Automatic Speech Recognition (ASR) dataset for the Oromo language, a widely spoken language in Ethiopia and neighboring regions. The dataset was collected through a crowd-sourcing initiative, encompassing a diverse range of speakers and phonetic variations. It consists of 100 hours of real-world audio recordings paired with transcriptions, covering read speech in both clean and noisy environments. This dataset addresses the critical need for ASR resources for the Oromo language which is underrepresented. To show its applicability for the ASR task, we conducted experiments using the Conformer model, achieving a Word Error Rate (WER) of 15.32% with hybrid CTC and AED loss and WER of 18.74% with pure CTC loss. Additionally, fine-tuning the Whisper model resulted in a significantly improved WER of 10.82%. These results establish baselines for Oromo ASR, highlighting both the challenges and the potential for improving ASR performance in Oromo. The dataset is publicly available at https://github.com/turinaf/sagalee and we encourage its use for further research and development in Oromo speech processing., Comment: Accepted for ICASSP2025 (2025 IEEE International Conference on Acoustics, Speech, and Signal Processing)

In this paper, Whisper, a large-scale pre-trained model for automatic speech recognition, is proposed to apply to speaker verification. A partial multi-scale feature aggregation (PMFA) approach is proposed based on a subset of Whisper encoder blocks to derive highly discriminative speaker embeddings.Experimental results demonstrate that using the middle to later blocks of the Whisper encoder keeps more speaker information. On the VoxCeleb1 and CN-Celeb1 datasets, our system achieves 1.42% and 8.23% equal error rates (EERs) respectively, receiving 0.58% and 1.81% absolute EER reductions over the ECAPA-TDNN baseline, and 0.46% and 0.97% over the ResNet34 baseline. Furthermore, our results indicate that using Whisper models trained on multilingual data can effectively enhance the model's robustness across languages. Finally, the low-rank adaptation approach is evaluated, which reduces the trainable model parameters by approximately 45 times while only slightly increasing EER by 0.2%., Comment: Accepted by Interspeech 2024

Automatic Speaker Verification (ASV) suffers from performance degradation in noisy conditions. To address this issue, we propose a novel adversarial learning framework that incorporates noise-disentanglement to establish a noise-independent speaker invariant embedding space. Specifically, the disentanglement module includes two encoders for separating speaker related and irrelevant information, respectively. The reconstruction module serves as a regularization term to constrain the noise. A feature-robust loss is also used to supervise the speaker encoder to learn noise-independent speaker embeddings without losing speaker information. In addition, adversarial training is introduced to discourage the speaker encoder from encoding acoustic condition information for achieving a speaker-invariant embedding space. Experiments on VoxCeleb1 indicate that the proposed method improves the performance of the speaker verification system under both clean and noisy conditions., Comment: 5 pages, accepted by Interspeech2024

End-to-end models have shown superior performance for automatic speech recognition (ASR). However, such models are often very large in size and thus challenging to deploy on resource-constrained edge devices. While quantisation can reduce model sizes, it can lead to increased word error rates (WERs). Although improved quantisation methods were proposed to address the issue of performance degradation, the fact that quantised models deployed on edge devices often target only on a small group of users is under-explored. To this end, we propose personalisation for quantised models (P4Q), a novel strategy that uses speaker adaptation (SA) to improve quantised end-to-end ASR models by fitting them to the characteristics of the target speakers. In this paper, we study the P4Q strategy based on Whisper and Conformer attention-based encoder-decoder (AED) end-to-end ASR models, which leverages a 4-bit block-wise NormalFloat4 (NF4) approach for quantisation and the low-rank adaptation (LoRA) approach for SA. Experimental results on the LibriSpeech and the TED-LIUM 3 corpora show that, with a 7-time reduction in model size and 1% extra speaker-specific parameters, 15.1% and 23.3% relative WER reductions were achieved on quantised Whisper and Conformer AED models respectively, comparing to the full precision models., Comment: submitted to ASRU 2023 Workshop

Mixture-of-experts (MoE) models have achieved excellent results in many tasks. However, conventional MoE models are often very large, making them challenging to deploy on resource-constrained edge devices. In this paper, we propose a novel speaker adaptive mixture of LoRA experts (SAML) approach, which uses low-rank adaptation (LoRA) modules as experts to reduce the number of trainable parameters in MoE. Specifically, SAML is applied to the quantised and personalised end-to-end automatic speech recognition models, which combines test-time speaker adaptation to improve the performance of heavily compressed models in speaker-specific scenarios. Experiments have been performed on the LibriSpeech and the TED-LIUM 3 corpora. Remarkably, with a 7x reduction in model size, 29.1% and 31.1% relative word error rate reductions were achieved on the quantised Whisper model and Conformer-based attention-based encoder-decoder ASR model respectively, comparing to the original full precision models., Comment: 5 pages, accepted by Interspeech 2024. arXiv admin note: substantial text overlap with arXiv:2309.09136

Recent end-to-end automatic speech recognition (ASR) models have become increasingly larger, making them particularly challenging to be deployed on resource-constrained devices. Model quantisation is an effective solution that sometimes causes the word error rate (WER) to increase. In this paper, a novel strategy of personalisation for a quantised model (PQM) is proposed, which combines speaker adaptive training (SAT) with model quantisation to improve the performance of heavily compressed models. Specifically, PQM uses a 4-bit NormalFloat Quantisation (NF4) approach for model quantisation and low-rank adaptation (LoRA) for SAT. Experiments have been performed on the LibriSpeech and the TED-LIUM 3 corpora. Remarkably, with a 7x reduction in model size and 1% additional speaker-specific parameters, 15.1% and 23.3% relative WER reductions were achieved on quantised Whisper and Conformer-based attention-based encoder-decoder ASR models respectively, comparing to the original full precision models., Comment: 5 pages, submitted to ICASSP 2024

The way that humans encode their emotion into speech signals is complex. For instance, an angry man may increase his pitch and speaking rate, and use impolite words. In this paper, we present a preliminary study on various emotional factors and investigate how each of them impacts modern emotion recognition systems. The key tool of our study is the SpeechFlow model presented recently, by which we are able to decompose speech signals into separate information factors (content, pitch, rhythm). Based on this decomposition, we carefully studied the performance of each information component and their combinations. We conducted the study on three different speech emotion corpora and chose an attention-based convolutional RNN as the emotion classifier. Our results show that rhythm is the most important component for emotional expression. Moreover, the cross-corpus results are very bad (even worse than guess), demonstrating that the present speech emotion recognition model is rather weak. Interestingly, by removing one or several unimportant components, the cross-corpus results can be improved. This demonstrates the potential of the decomposition approach towards a generalizable emotion recognition.

The choice of an optimal time-frequency resolution is usually a difficult but important step in tasks involving speech signal classification, e.g., speech anti-spoofing. The variations of the performance with different choices of timefrequency resolutions can be as large as those with different model architectures, which makes it difficult to judge what the improvement actually comes from when a new network architecture is invented and introduced as the classifier. In this paper, we propose a multi-resolution front-end for feature extraction in an end-to-end classification framework. Optimal weighted combinations of multiple time-frequency resolutions will be learned automatically given the objective of a classification task. Features extracted with different time-frequency resolutions are weighted and concatenated as inputs to the successive networks, where the weights are predicted by a learnable neural network inspired by the weighting block in squeeze-and-excitation networks (SENet). Furthermore, the refinement of the chosen timefrequency resolutions is investigated by pruning the ones with relatively low importance, which reduces the complexity and size of the model. The proposed method is evaluated on the tasks of speech anti-spoofing in ASVSpoof 2019 and its superiority has been justified by comparing with similar baselines., Comment: submitted to ICASSP 2022

In authentication scenarios, applications of practical speaker verification systems usually require a person to read a dynamic authentication text. Previous studies played an audio adversarial example as a digital signal to perform physical attacks, which would be easily rejected by audio replay detection modules. This work shows that by playing our crafted adversarial perturbation as a separate source when the adversary is speaking, the practical speaker verification system will misjudge the adversary as a target speaker. A two-step algorithm is proposed to optimize the universal adversarial perturbation to be text-independent and has little effect on the authentication text recognition. We also estimated room impulse response (RIR) in the algorithm which allowed the perturbation to be effective after being played over the air. In the physical experiment, we achieved targeted attacks with success rate of 100%, while the word error rate (WER) on speech recognition was only increased by 3.55%. And recorded audios could pass replay detection for the live person speaking., Comment: 6 pages, 2 figures

Research on speaker recognition is extending to address the vulnerability in the wild conditions, among which genre mismatch is perhaps the most challenging, for instance, enrollment with reading speech while testing with conversational or singing audio. This mismatch leads to complex and composite inter-session variations, both intrinsic (i.e., speaking style, physiological status) and extrinsic (i.e., recording device, background noise). Unfortunately, the few existing multi-genre corpora are not only limited in size but are also recorded under controlled conditions, which cannot support conclusive research on the multi-genre problem. In this work, we firstly publish CN-Celeb, a large-scale multi-genre corpus that includes in-the-wild speech utterances of 3,000 speakers in 11 different genres. Secondly, using this dataset, we conduct a comprehensive study on the multi-genre phenomenon, in particular the impact of the multi-genre challenge on speaker recognition and the performance gain when the new dataset is used to conduct multi-genre training., Comment: submitted to Speech Communication

Domain mismatch often occurs in real applications and causes serious performance reduction on speaker verification systems. The common wisdom is to collect cross-domain data and train a multi-domain PLDA model, with the hope to learn a domain-independent speaker subspace. In this paper, we firstly present an empirical study to show that simply adding cross-domain data does not help performance in conditions with enrollment-test mismatch. Careful analysis shows that this striking result is caused by the incoherent statistics between the enrollment and test conditions. Based on this analysis, we present a decoupled scoring approach that can maximally squeeze the value of cross-domain labels and obtain optimal verification scores when the enrollment and test are mismatched. When the statistics are coherent, the new formulation falls back to the conventional PLDA. Experimental results on cross-channel test show that the proposed approach is highly effective and is a principle solution to domain mismatch., Comment: Submitted to ICASSP 2021

Various information factors are blended in speech signals, which forms the primary difficulty for most speech information processing tasks. An intuitive idea is to factorize speech signal into individual information factors (e.g., phonetic content and speaker trait), though it turns out to be highly challenging. This paper presents a speech factorization approach based on a novel factorial discriminative normalization flow model (factorial DNF). Experiments conducted on a two-factor case that involves phonetic content and speaker trait demonstrates that the proposed factorial DNF has powerful capability to factorize speech signals and outperforms several comparative models in terms of information representation and manipulation., Comment: Submitted to ICASSP 2021

The technique of transforming voices in order to hide the real identity of a speaker is called voice disguise, among which automatic voice disguise (AVD) by modifying the spectral and temporal characteristics of voices with miscellaneous algorithms are easily conducted with softwares accessible to the public. AVD has posed great threat to both human listening and automatic speaker verification (ASV). In this paper, we have found that ASV is not only a victim of AVD but could be a tool to beat some simple types of AVD. Firstly, three types of AVD, pitch scaling, vocal tract length normalization (VTLN) and voice conversion (VC), are introduced as representative methods. State-of-the-art ASV methods are subsequently utilized to objectively evaluate the impact of AVD on ASV by equal error rates (EER). Moreover, an approach to restore disguised voice to its original version is proposed by minimizing a function of ASV scores w.r.t. restoration parameters. Experiments are then conducted on disguised voices from Voxceleb, a dataset recorded in real-world noisy scenario. The results have shown that, for the voice disguise by pitch scaling, the proposed approach obtains an EER around 7% comparing to the 30% EER of a recently proposed baseline using the ratio of fundamental frequencies. The proposed approach generalizes well to restore the disguise with nonlinear frequency warping in VTLN by reducing its EER from 34.3% to 18.5%. However, it is difficult to restore the source speakers in VC by our approach, where more complex forms of restoration functions or other paralinguistic cues might be necessary to restore the nonlinear transform in VC. Finally, contrastive visualization on ASV features with and without restoration illustrate the role of the proposed approach in an intuitive way., Comment: accepted for publication

Various informative factors mixed in speech signals, leading to great difficulty when decoding any of the factors. An intuitive idea is to factorize each speech frame into individual informative factors, though it turns out to be highly difficult. Recently, we found that speaker traits, which were assumed to be long-term distributional properties, are actually short-time patterns, and can be learned by a carefully designed deep neural network (DNN). This discovery motivated a cascade deep factorization (CDF) framework that will be presented in this paper. The proposed framework infers speech factors in a sequential way, where factors previously inferred are used as conditional variables when inferring other factors. We will show that this approach can effectively factorize speech signals, and using these factors, the original speech spectrum can be recovered with a high accuracy. This factorization and reconstruction approach provides potential values for many speech processing tasks, e.g., speaker recognition and emotion recognition, as will be demonstrated in the paper., Comment: Accepted by ICASSP 2018. arXiv admin note: substantial text overlap with arXiv:1706.01777

In recent studies, it has shown that speaker patterns can be learned from very short speech segments (e.g., 0.3 seconds) by a carefully designed convolutional & time-delay deep neural network (CT-DNN) model. By enforcing the model to discriminate the speakers in the training data, frame-level speaker features can be derived from the last hidden layer. In spite of its good performance, a potential problem of the present model is that it involves a parametric classifier, i.e., the last affine layer, which may consume some discriminative knowledge, thus leading to `information leak' for the feature learning. This paper presents a full-info training approach that discards the parametric classifier and enforces all the discriminative knowledge learned by the feature net. Our experiments on the Fisher database demonstrate that this new training scheme can produce more coherent features, leading to consistent and notable performance improvement on the speaker verification task., Comment: Accepted by ICASSP 2018

End-to-end learning treats the entire system as a whole adaptable black box, which, if sufficient data are available, may learn a system that works very well for the target task. This principle has recently been applied to several prototype research on speaker verification (SV), where the feature learning and classifier are learned together with an objective function that is consistent with the evaluation metric. An opposite approach to end-to-end is feature learning, which firstly trains a feature learning model, and then constructs a back-end classifier separately to perform SV. Recently, both approaches achieved significant performance gains on SV, mainly attributed to the smart utilization of deep neural networks. However, the two approaches have not been carefully compared, and their respective advantages have not been well discussed. In this paper, we compare the end-to-end and feature learning approaches on a text-independent SV task. Our experiments on a dataset sampled from the Fisher database and involving 5,000 speakers demonstrated that the feature learning approach outperformed the end-to-end approach. This is a strong support for the feature learning approach, at least with data and computation resources similar to ours.

Existing speaker verification (SV) systems often suffer from performance degradation if there is any language mismatch between model training, speaker enrollment, and test. A major cause of this degradation is that most existing SV methods rely on a probabilistic model to infer the speaker factor, so any significant change on the distribution of the speech signal will impact the inference. Recently, we proposed a deep learning model that can learn how to extract the speaker factor by a deep neural network (DNN). By this feature learning, an SV system can be constructed with a very simple back-end model. In this paper, we investigate the robustness of the feature-based SV system in situations with language mismatch. Our experiments were conducted on a complex cross-lingual scenario, where the model training was in English, and the enrollment and test were in Chinese or Uyghur. The experiments demonstrated that the feature-based system outperformed the i-vector system with a large margin, particularly with language mismatch between enrollment and test.

For practical automatic speaker verification (ASV) systems, replay attack poses a true risk. By replaying a pre-recorded speech signal of the genuine speaker, ASV systems tend to be easily fooled. An effective replay detection method is therefore highly desirable. In this study, we investigate a major difficulty in replay detection: the over-fitting problem caused by variability factors in speech signal. An F-ratio probing tool is proposed and three variability factors are investigated using this tool: speaker identity, speech content and playback & recording device. The analysis shows that device is the most influential factor that contributes the highest over-fitting risk. A frequency warping approach is studied to alleviate the over-fitting problem, as verified on the ASV-spoof 2017 database.

Decision making is an important component in a speaker verification system. For the conventional GMM-UBM architecture, the decision is usually conducted based on the log likelihood ratio of the test utterance against the GMM of the claimed speaker and the UBM. This single-score decision is simple but tends to be sensitive to the complex variations in speech signals (e.g. text content, channel, speaking style, etc.). In this paper, we propose a decision making approach based on multiple scores derived from a set of cohort GMMs (cohort scores). Importantly, these cohort scores are not simply averaged as in conventional cohort methods; instead, we employ a powerful discriminative model as the decision maker. Experimental results show that the proposed method delivers substantial performance improvement over the baseline system, especially when a deep neural network (DNN) is used as the decision maker, and the DNN input involves some statistical features derived from the cohort scores., Comment: APSIPA ASC 2016

Probabilistic linear discriminant analysis (PLDA) is a popular normalization approach for the i-vector model, and has delivered state-of-the-art performance in speaker recognition. A potential problem of the PLDA model, however, is that it essentially assumes Gaussian distributions over speaker vectors, which is not always true in practice. Additionally, the objective function is not directly related to the goal of the task, e.g., discriminating true speakers and imposters. In this paper, we propose a max-margin metric learning approach to solve the problems. It learns a linear transform with a criterion that the margin between target and imposter trials are maximized. Experiments conducted on the SRE08 core test show that compared to PLDA, the new approach can obtain comparable or even better performance, though the scoring is simply a cosine computation.

The popular i-vector model represents speakers as low-dimensional continuous vectors (i-vectors), and hence it is a way of continuous speaker embedding. In this paper, we investigate binary speaker embedding, which transforms i-vectors to binary vectors (codes) by a hash function. We start from locality sensitive hashing (LSH), a simple binarization approach where binary codes are derived from a set of random hash functions. A potential problem of LSH is that the randomly sampled hash functions might be suboptimal. We therefore propose an improved Hamming distance learning approach, where the hash function is learned by a variable-sized block training that projects each dimension of the original i-vectors to variable-sized binary codes independently. Our experiments show that binary speaker embedding can deliver competitive or even better results on both speaker verification and identification tasks, while the memory usage and the computation cost are significantly reduced.