Multi-head attention has each of the attention heads collect salient information from different parts of an input sequence, making it a powerful mechanism for sequence modeling. Multilingual and multi-domain learning are common scenarios for sequence modeling, where the key challenge is to maximize positive transfer and mitigate negative transfer across languages and domains. In this paper, we find that non-selective attention sharing is sub-optimal for achieving good generalization across all languages and domains. We further propose attention sharing strategies to facilitate parameter sharing and specialization in multilingual and multi-domain sequence modeling. Our approach automatically learns shared and specialized attention heads for different languages and domains to mitigate their interference. Evaluated in various tasks including speech recognition, text-to-text and speech-to-text translation, the proposed attention sharing strategies consistently bring gains to sequence models built upon multi-head attention. For speech-to-text translation, our approach yields an average of $+2.0$ BLEU over $13$ language directions in multilingual setting and $+2.0$ BLEU over $3$ domains in multi-domain setting.
With the increase in the availability of speech from varied domains, it is imperative to use such out-of-domain data to improve existing speech systems. Domain adaptation is a prominent pre-processing approach for this. We investigate it for adapt microphone speech to the telephone domain. Specifically, we explore CycleGAN-based unpaired translation of microphone data to improve the x-vector/speaker embedding network for Telephony Speaker Verification. We first demonstrate the efficacy of this on real challenging data and then, to improve further, we modify the CycleGAN formulation to make the adaptation task-specific. We modify CycleGAN's identity loss, cycle-consistency loss, and adversarial loss to operate in the deep feature space. Deep features of a signal are extracted from an auxiliary (speaker embedding) network and, hence, preserves speaker identity. Our 3D convolution-based Deep Feature Discriminators (DFD) show relative improvements of 5-10% in terms of equal error rate. To dive deeper, we study a challenging scenario of pooling (adapted) microphone and telephone data with data augmentations and telephone codecs. Finally, we highlight the sensitivity of CycleGAN hyper-parameters and introduce a parameter called probability of adaptation.
Most existing sequence labelling models rely on a fixed decomposition of a target sequence into a sequence of basic units. These methods suffer from two major drawbacks: 1) the set of basic units is fixed, such as the set of words, characters or phonemes in speech recognition, and 2) the decomposition of target sequences is fixed. These drawbacks usually result in sub-optimal performance of modeling sequences. In this pa- per, we extend the popular CTC loss criterion to alleviate these limitations, and propose a new loss function called Gram-CTC. While preserving the advantages of CTC, Gram-CTC automatically learns the best set of basic units (grams), as well as the most suitable decomposition of tar- get sequences. Unlike CTC, Gram-CTC allows the model to output variable number of characters at each time step, which enables the model to capture longer term dependency and improves the computational efficiency. We demonstrate that the proposed Gram-CTC improves CTC in terms of both performance and efficiency on the large vocabulary speech recognition task at multiple scales of data, and that with Gram-CTC we can outperform the state-of-the-art on a standard speech benchmark.
Most of the speech recognition systems recover only words in the speech and fail to capture emotions. Users have to manually add emoji(s) in text for adding tone and making communication fun. Though there is much work done on punctuation addition on transcribed speech, the area of emotion addition is untouched. In this paper, we propose a novel on-device pipeline to enrich the voice input experience. It involves, given a blob of transcribed text, intelligently processing and identifying structure where emoji insertion makes sense. Moreover, it includes semantic text analysis to predict emoji for each of the sub-parts for which we propose a novel architecture Attention-based Char Aware (ACA) LSTM which handles Out-Of-Vocabulary (OOV) words as well. All these tasks are executed completely on-device and hence can aid on-device dictation systems. To the best of our knowledge, this is the first work that shows how to add emoji(s) in the transcribed text. We demonstrate that our components achieve comparable results to previous neural approaches for punctuation addition and emoji prediction with 80% fewer parameters. Overall, our proposed model has a very small memory footprint of a mere 4MB to suit on-device deployment.
End-to-end approaches for sequence tasks are becoming increasingly popular. Yet for complex sequence tasks, like speech translation, systems that cascade several models trained on sub-tasks have shown to be superior, suggesting that the compositionality of cascaded systems simplifies learning and enables sophisticated search capabilities. In this work, we present an end-to-end framework that exploits compositionality to learn searchable hidden representations at intermediate stages of a sequence model using decomposed sub-tasks. These hidden intermediates can be improved using beam search to enhance the overall performance and can also incorporate external models at intermediate stages of the network to re-score or adapt towards out-of-domain data. One instance of the proposed framework is a Multi-Decoder model for speech translation that extracts the searchable hidden intermediates from a speech recognition sub-task. The model demonstrates the aforementioned benefits and outperforms the previous state-of-the-art by around +6 and +3 BLEU on the two test sets of Fisher-CallHome and by around +3 and +4 BLEU on the English-German and English-French test sets of MuST-C.
We propose a cloud-based multimodal dialog platform for the remote assessment and monitoring of Amyotrophic Lateral Sclerosis (ALS) at scale. This paper presents our vision, technology setup, and an initial investigation of the efficacy of the various acoustic and visual speech metrics automatically extracted by the platform. 82 healthy controls and 54 people with ALS (pALS) were instructed to interact with the platform and completed a battery of speaking tasks designed to probe the acoustic, articulatory, phonatory, and respiratory aspects of their speech. We find that multiple acoustic (rate, duration, voicing) and visual (higher order statistics of the jaw and lip) speech metrics show statistically significant differences between controls, bulbar symptomatic and bulbar pre-symptomatic patients. We report on the sensitivity and specificity of these metrics using five-fold cross-validation. We further conducted a LASSO-LARS regression analysis to uncover the relative contributions of various acoustic and visual features in predicting the severity of patients' ALS (as measured by their self-reported ALSFRS-R scores). Our results provide encouraging evidence of the utility of automatically extracted audiovisual analytics for scalable remote patient assessment and monitoring in ALS.
Alzheimer's Disease (AD) is nowadays the most common form of dementia, and its automatic detection can help to identify symptoms at early stages, so that preventive actions can be carried out. Moreover, non-intrusive techniques based on spoken data are crucial for the development of AD automatic detection systems. In this light, this paper is presented as a contribution to the ADReSS Challenge, aiming at improving AD automatic detection from spontaneous speech. To this end, recordings from 108 participants, which are age-, gender-, and AD condition-balanced, have been used as training set to perform two different tasks: classification into AD/non-AD conditions, and regression over the Mini-Mental State Examination (MMSE) scores. Both tasks have been performed extracting 28 features from speech -- based on prosody and voice quality -- and 51 features from the transcriptions -- based on lexical and turn-taking information. Our results achieved up to 87.5 % of classification accuracy using a Random Forest classifier, and 4.54 of RMSE using a linear regression with stochastic gradient descent over the provided test set. This shows promising results in the automatic detection of Alzheimer's Disease through speech and lexical features.
We present the Geometric-Wave Acoustic (GWA) dataset, a large-scale audio dataset of over 2 million synthetic room impulse responses (IRs) and their corresponding detailed geometric and simulation configurations. Our dataset samples acoustic environments from over 6.8K high-quality diverse and professionally designed houses represented as semantically labeled 3D meshes. We also present a novel real-world acoustic materials assignment scheme based on semantic matching that uses a sentence transformer model. We compute high-quality impulse responses corresponding to accurate low-frequency and high-frequency wave effects by automatically calibrating geometric acoustic ray-tracing with a finite-difference time-domain wave solver. We demonstrate the higher accuracy of our IRs by comparing with recorded IRs from complex real-world environments. The code and the full dataset will be released at the time of publication. Moreover, we highlight the benefits of GWA on audio deep learning tasks such as automated speech recognition, speech enhancement, and speech separation. We observe significant improvement over prior synthetic IR datasets in all tasks due to using our dataset.
This work addresses the mismatch problem between the distribution of training data (source) and testing data (target), in the challenging context of dysarthric speech recognition. We focus on Speaker Adaptation (SA) in command speech recognition, where data from multiple sources (i.e., multiple speakers) are available. Specifically, we propose an unsupervised Multi-Source Domain Adaptation (MSDA) algorithm based on optimal-transport, called MSDA via Weighted Joint Optimal Transport (MSDA-WJDOT). We achieve a Command Error Rate relative reduction of 16% and 7% over the speaker-independent model and the best competitor method, respectively. The strength of the proposed approach is that, differently from any other existing SA method, it offers an interpretable model that can also be exploited, in this context, to diagnose dysarthria without any specific training. Indeed, it provides a closeness measure between the target and the source speakers, reflecting their similarity in terms of speech characteristics. Based on the similarity between the target speaker and the healthy/dysarthric source speakers, we then define the healthy/dysarthric score of the target speaker that we leverage to perform dysarthria detection. This approach does not require any additional training and achieves a 95% accuracy in the dysarthria diagnosis.
In this paper, we propose an innovative approach to perform speaker recognition by fusing two recently introduced deep neural networks (DNNs) namely - SincNet and X-Vector. The idea behind using SincNet filters on the raw speech waveform is to extract more distinguishing frequency-related features in the initial convolution layers of the CNN architecture. X-Vectors are used to take advantage of the fact that this embedding is an efficient method to churn out fixed dimension features from variable length speech utterances, something which is challenging in plain CNN techniques, making it efficient both in terms of speed and accuracy. Our approach uses the best of both worlds by combining X-vector in the later layers while using SincNet filters in the initial layers of our deep model. This approach allows the network to learn better embedding and converge quicker. Previous works use either X-Vector or SincNet Filters or some modifications, however we introduce a novel fusion architecture wherein we have combined both the techniques to gather more information about the speech signal hence, giving us better results. Our method focuses on the VoxCeleb1 dataset for speaker recognition, and we have used it for both training and testing purposes.