Sameer Khurana

Sameer Khurana, Antoine Laurent, James Glass · mit

We propose the SAMU-XLSR: Semantically-Aligned Multimodal Utterance-level Cross-Lingual Speech Representation learning framework. Unlike previous works on speech representation learning, which learns multilingual contextual speech embedding at the resolution of an acoustic frame (10-20ms), this work focuses on learning multimodal (speech-text) multilingual speech embedding at the resolution of a sentence (5-10s) such that the embedding vector space is semantically aligned across different languages. We combine state-of-the-art multilingual acoustic frame-level speech representation learning model XLS-R with the Language Agnostic BERT Sentence Embedding (LaBSE) model to create an utterance-level multimodal multilingual speech encoder SAMU-XLSR. Although we train SAMU-XLSR with only multilingual transcribed speech data, cross-lingual speech-text and speech-speech associations emerge in its learned representation space. To substantiate our claims, we use SAMU-XLSR speech encoder in combination with a pre-trained LaBSE text sentence encoder for cross-lingual speech-to-text translation retrieval, and SAMU-XLSR alone for cross-lingual speech-to-speech translation retrieval. We highlight these applications by performing several cross-lingual text and speech translation retrieval tasks across several datasets.

Yuan Gong, Sameer Khurana, Andrew Rouditchenko et al. · mit

Audio classification is an active research area with a wide range of applications. Over the past decade, convolutional neural networks (CNNs) have been the de-facto standard building block for end-to-end audio classification models. Recently, neural networks based solely on self-attention mechanisms such as the Audio Spectrogram Transformer (AST) have been shown to outperform CNNs. In this paper, we find an intriguing interaction between the two very different models - CNN and AST models are good teachers for each other. When we use either of them as the teacher and train the other model as the student via knowledge distillation (KD), the performance of the student model noticeably improves, and in many cases, is better than the teacher model. In our experiments with this CNN/Transformer Cross-Model Knowledge Distillation (CMKD) method we achieve new state-of-the-art performance on FSD50K, AudioSet, and ESC-50.

Nauman Dawalatabad, Sameer Khurana, Antoine Laurent et al. · mit

Pseudo-label (PL) filtering forms a crucial part of Self-Training (ST) methods for unsupervised domain adaptation. Dropout-based Uncertainty-driven Self-Training (DUST) proceeds by first training a teacher model on source domain labeled data. Then, the teacher model is used to provide PLs for the unlabeled target domain data. Finally, we train a student on augmented labeled and pseudo-labeled data. The process is iterative, where the student becomes the teacher for the next DUST iteration. A crucial step that precedes the student model training in each DUST iteration is filtering out noisy PLs that could lead the student model astray. In DUST, we proposed a simple, effective, and theoretically sound PL filtering strategy based on the teacher model's uncertainty about its predictions on unlabeled speech utterances. We estimate the model's uncertainty by computing disagreement amongst multiple samples drawn from the teacher model during inference by injecting noise via dropout. In this work, we show that DUST's PL filtering, as initially used, may fail under severe source and target domain mismatch. We suggest several approaches to eliminate or alleviate this issue. Further, we bring insights from the research in neural network model calibration to DUST and show that a well-calibrated model correlates strongly with a positive outcome of the DUST PL filtering step.

Sameer Khurana, Nauman Dawalatabad, Antoine Laurent et al. · mit

Research in multilingual speech-to-text translation is topical. Having a single model that supports multiple translation tasks is desirable. The goal of this work it to improve cross-lingual transfer learning in multilingual speech-to-text translation via semantic knowledge distillation. We show that by initializing the encoder of the encoder-decoder sequence-to-sequence translation model with SAMU-XLS-R, a multilingual speech transformer encoder trained using multi-modal (speech-text) semantic knowledge distillation, we achieve significantly better cross-lingual task knowledge transfer than the baseline XLS-R, a multilingual speech transformer encoder trained via self-supervised learning. We demonstrate the effectiveness of our approach on two popular datasets, namely, CoVoST-2 and Europarl. On the 21 translation tasks of the CoVoST-2 benchmark, we achieve an average improvement of 12.8 BLEU points over the baselines. In the zero-shot translation scenario, we achieve an average gain of 18.8 and 11.9 average BLEU points on unseen medium and low-resource languages. We make similar observations on Europarl speech translation benchmark.

Victoria Mingote, Pablo Gimeno, Luis Vicente et al. · mit

This paper proposes a direct text to speech translation system using discrete acoustic units. This framework employs text in different source languages as input to generate speech in the target language without the need for text transcriptions in this language. Motivated by the success of acoustic units in previous works for direct speech to speech translation systems, we use the same pipeline to extract the acoustic units using a speech encoder combined with a clustering algorithm. Once units are obtained, an encoder-decoder architecture is trained to predict them. Then a vocoder generates speech from units. Our approach for direct text to speech translation was tested on the new CVSS corpus with two different text mBART models employed as initialisation. The systems presented report competitive performance for most of the language pairs evaluated. Besides, results show a remarkable improvement when initialising our proposed architecture with a model pre-trained with more languages.

Antonio Almudévar, José Miguel Hernández-Lobato, Sameer Khurana et al. · mit

Contrastive losses have been extensively used as a tool for multimodal representation learning. However, it has been empirically observed that their use is not effective to learn an aligned representation space. In this paper, we argue that this phenomenon is caused by the presence of modality-specific information in the representation space. Although some of the most widely used contrastive losses maximize the mutual information between representations of both modalities, they are not designed to remove the modality-specific information. We give a theoretical description of this problem through the lens of the Information Bottleneck Principle. We also empirically analyze how different hyperparameters affect the emergence of this phenomenon in a controlled experimental setup. Finally, we propose a regularization term in the loss function that is derived by means of a variational approximation and aims to increase the representational alignment. We analyze in a set of controlled experiments and real-world applications the advantages of including this regularization term.

Sameer Khurana, Dominik Klement, Antoine Laurent et al.

We propose Hierarchical Audio Codec (HAC), a unified neural speech codec that factorizes its bottleneck into three linguistic levels-acoustic, phonetic, and lexical-within a single model. HAC leverages two knowledge distillation objectives: one from a pre-trained speech encoder (HuBERT) for phoneme-level structure, and another from a text-based encoder (LaBSE) for lexical cues. Experiments on English and multilingual data show that HAC's factorized bottleneck yields disentangled token sets: one aligns with phonemes, while another captures word-level semantics. Quantitative evaluations confirm that HAC tokens preserve naturalness and provide interpretable linguistic information, outperforming single-level baselines in both disentanglement and reconstruction quality. These findings underscore HAC's potential as a unified discrete speech representation, bridging acoustic detail and lexical meaning for downstream speech generation and understanding tasks.

Santiago Cuervo, Adel Moumen, Yanis Labrak et al. · mit

Text-Speech Language Models (TSLMs) -- language models trained to jointly process and generate text and speech -- are commonly trained through an early modality fusion/fission approach, in which both modalities are fed and predicted from a shared backbone via linear layers. We hypothesize that this approach limits cross-modal transfer by neglecting feature compositionality -- specifically, the finer-grained nature of speech representations compared to text -- preventing the emergence of a shared feature hierarchy within model layers. In this paper, we argue that this limitation can be addressed through late fusion and fission, with a fission process that accesses both high- and low-level features for speech generation. Our models implementing these principles, SmolTolk, rival or surpass state-of-the-art TSLMs trained with orders of magnitude more compute, and achieve significantly improved cross-modal performance relative to early fusion/fission baselines. Representation analyses further suggest that our method enhances the model's ability to abstract higher-level, more semantic features from speech, and leads to increasingly shared representation spaces across layers.

Andrew Rouditchenko, Sameer Khurana, Samuel Thomas et al.

Recent models such as XLS-R and Whisper have made multilingual speech technologies more accessible by pre-training on audio from around 100 spoken languages each. However, there are thousands of spoken languages worldwide, and adapting to new languages is an important problem. In this work, we aim to understand which model adapts better to languages unseen during pre-training. We fine-tune both models on 13 unseen languages and 18 seen languages. Our results show that the number of hours seen per language and language family during pre-training is predictive of how the models compare, despite the significant differences in the pre-training methods.

Sameer Khurana, Antoine Laurent, James Glass

We propose a simple and effective cross-lingual transfer learning method to adapt monolingual wav2vec-2.0 models for Automatic Speech Recognition (ASR) in resource-scarce languages. We show that a monolingual wav2vec-2.0 is a good few-shot ASR learner in several languages. We improve its performance further via several iterations of Dropout Uncertainty-Driven Self-Training (DUST) by using a moderate-sized unlabeled speech dataset in the target language. A key finding of this work is that the adapted monolingual wav2vec-2.0 achieves similar performance as the topline multilingual XLSR model, which is trained on fifty-three languages, on the target language ASR task.

Cheng-I Jeff Lai, Yang Zhang, Alexander H. Liu et al.

Self-supervised speech representation learning (speech SSL) has demonstrated the benefit of scale in learning rich representations for Automatic Speech Recognition (ASR) with limited paired data, such as wav2vec 2.0. We investigate the existence of sparse subnetworks in pre-trained speech SSL models that achieve even better low-resource ASR results. However, directly applying widely adopted pruning methods such as the Lottery Ticket Hypothesis (LTH) is suboptimal in the computational cost needed. Moreover, we show that the discovered subnetworks yield minimal performance gain compared to the original dense network. We present Prune-Adjust-Re-Prune (PARP), which discovers and finetunes subnetworks for much better performance, while only requiring a single downstream ASR finetuning run. PARP is inspired by our surprising observation that subnetworks pruned for pre-training tasks need merely a slight adjustment to achieve a sizeable performance boost in downstream ASR tasks. Extensive experiments on low-resource ASR verify (1) sparse subnetworks exist in mono-lingual/multi-lingual pre-trained speech SSL, and (2) the computational advantage and performance gain of PARP over baseline pruning methods. In particular, on the 10min Librispeech split without LM decoding, PARP discovers subnetworks from wav2vec 2.0 with an absolute 10.9%/12.6% WER decrease compared to the full model. We further demonstrate the effectiveness of PARP via: cross-lingual pruning without any phone recognition degradation, the discovery of a multi-lingual subnetwork for 10 spoken languages in 1 finetuning run, and its applicability to pre-trained BERT/XLNet for natural language tasks.

Sameer Khurana, Niko Moritz, Takaaki Hori et al.

The performance of automatic speech recognition (ASR) systems typically degrades significantly when the training and test data domains are mismatched. In this paper, we show that self-training (ST) combined with an uncertainty-based pseudo-label filtering approach can be effectively used for domain adaptation. We propose DUST, a dropout-based uncertainty-driven self-training technique which uses agreement between multiple predictions of an ASR system obtained for different dropout settings to measure the model's uncertainty about its prediction. DUST excludes pseudo-labeled data with high uncertainties from the training, which leads to substantially improved ASR results compared to ST without filtering, and accelerates the training time due to a reduced training data set. Domain adaptation experiments using WSJ as a source domain and TED-LIUM 3 as well as SWITCHBOARD as the target domains show that up to 80% of the performance of a system trained on ground-truth data can be recovered.

Sameer Khurana, Antoine Laurent, James Glass

More than half of the 7,000 languages in the world are in imminent danger of going extinct. Traditional methods of documenting language proceed by collecting audio data followed by manual annotation by trained linguists at different levels of granularity. This time consuming and painstaking process could benefit from machine learning. Many endangered languages do not have any orthographic form but usually have speakers that are bi-lingual and trained in a high resource language. It is relatively easy to obtain textual translations corresponding to speech. In this work, we provide a multimodal machine learning framework for speech representation learning by exploiting the correlations between the two modalities namely speech and its corresponding text translation. Here, we construct a convolutional neural network audio encoder capable of extracting linguistic representations from speech. The audio encoder is trained to perform a speech-translation retrieval task in a contrastive learning framework. By evaluating the learned representations on a phone recognition task, we demonstrate that linguistic representations emerge in the audio encoder's internal representations as a by-product of learning to perform the retrieval task.

Sameer Khurana, Antoine Laurent, Wei-Ning Hsu et al.

Probabilistic Latent Variable Models (LVMs) provide an alternative to self-supervised learning approaches for linguistic representation learning from speech. LVMs admit an intuitive probabilistic interpretation where the latent structure shapes the information extracted from the signal. Even though LVMs have recently seen a renewed interest due to the introduction of Variational Autoencoders (VAEs), their use for speech representation learning remains largely unexplored. In this work, we propose Convolutional Deep Markov Model (ConvDMM), a Gaussian state-space model with non-linear emission and transition functions modelled by deep neural networks. This unsupervised model is trained using black box variational inference. A deep convolutional neural network is used as an inference network for structured variational approximation. When trained on a large scale speech dataset (LibriSpeech), ConvDMM produces features that significantly outperform multiple self-supervised feature extracting methods on linear phone classification and recognition on the Wall Street Journal dataset. Furthermore, we found that ConvDMM complements self-supervised methods like Wav2Vec and PASE, improving on the results achieved with any of the methods alone. Lastly, we find that ConvDMM features enable learning better phone recognizers than any other features in an extreme low-resource regime with few labeled training examples.

Adrian Łańcucki, Jan Chorowski, Guillaume Sanchez et al.

In this paper we demonstrate methods for reliable and efficient training of discrete representation using Vector-Quantized Variational Auto-Encoder models (VQ-VAEs). Discrete latent variable models have been shown to learn nontrivial representations of speech, applicable to unsupervised voice conversion and reaching state-of-the-art performance on unit discovery tasks. For unsupervised representation learning, they became viable alternatives to continuous latent variable models such as the Variational Auto-Encoder (VAE). However, training deep discrete variable models is challenging, due to the inherent non-differentiability of the discretization operation. In this paper we focus on VQ-VAE, a state-of-the-art discrete bottleneck model shown to perform on par with its continuous counterparts. It quantizes encoder outputs with on-line $k$-means clustering. We show that the codebook learning can suffer from poor initialization and non-stationarity of clustered encoder outputs. We demonstrate that these can be successfully overcome by increasing the learning rate for the codebook and periodic date-dependent codeword re-initialization. As a result, we achieve more robust training across different tasks, and significantly increase the usage of latent codewords even for large codebooks. This has practical benefit, for instance, in unsupervised representation learning, where large codebooks may lead to disentanglement of latent representations.

Sameer Khurana, Ahmed Ali, James Glass

We present the speech to text transcription system, called DARTS, for low resource Egyptian Arabic dialect. We analyze the following; transfer learning from high resource broadcast domain to low-resource dialectal domain and semi-supervised learning where we use in-domain unlabeled audio data collected from YouTube. Key features of our system are: A deep neural network acoustic model that consists of a front end Convolutional Neural Network (CNN) followed by several layers of Time Delayed Neural Network (TDNN) and Long-Short Term Memory Recurrent Neural Network (LSTM); sequence discriminative training of the acoustic model; n-gram and recurrent neural network language model for decoding and N-best list rescoring. We show that a simple transfer learning method can achieve good results. The results are further improved by using unlabeled data from YouTube in a semi-supervised setup. Various systems are combined to give the final system that achieves the lowest word error on on the community standard Egyptian-Arabic speech dataset (MGB-3).

Sameer Khurana, Ahmed Ali, Steve Renals

In this work, we present a new Vector Space Model (VSM) of speech utterances for the task of spoken dialect identification. Generally, DID systems are built using two sets of features that are extracted from speech utterances; acoustic and phonetic. The acoustic and phonetic features are used to form vector representations of speech utterances in an attempt to encode information about the spoken dialects. The Phonotactic and Acoustic VSMs, thus formed, are used for the task of DID. The aim of this paper is to construct a single VSM that encodes information about spoken dialects from both the Phonotactic and Acoustic VSMs. Given the two views of the data, we make use of a well known multi-view dimensionality reduction technique known as Canonical Correlation Analysis (CCA), to form a single vector representation for each speech utterance that encodes dialect specific discriminative information from both the phonetic and acoustic representations. We refer to this approach as feature space combination approach and show that our CCA based feature vector representation performs better on the Arabic DID task than the phonetic and acoustic feature representations used alone. We also present the feature space combination approach as a viable alternative to the model based combination approach, where two DID systems are built using the two VSMs (Phonotactic and Acoustic) and the final prediction score is the output score combination from the two systems.

Ahmed Ali, Najim Dehak, Patrick Cardinal et al.

We investigate different approaches for dialect identification in Arabic broadcast speech, using phonetic, lexical features obtained from a speech recognition system, and acoustic features using the i-vector framework. We studied both generative and discriminate classifiers, and we combined these features using a multi-class Support Vector Machine (SVM). We validated our results on an Arabic/English language identification task, with an accuracy of 100%. We used these features in a binary classifier to discriminate between Modern Standard Arabic (MSA) and Dialectal Arabic, with an accuracy of 100%. We further report results using the proposed method to discriminate between the five most widely used dialects of Arabic: namely Egyptian, Gulf, Levantine, North African, and MSA, with an accuracy of 52%. We discuss dialect identification errors in the context of dialect code-switching between Dialectal Arabic and MSA, and compare the error pattern between manually labeled data, and the output from our classifier. We also release the train and test data as standard corpus for dialect identification.