In this contribution, we investigate the effectiveness of deep fusion of text and audio features for categorical and dimensional speech emotion recognition (SER). We propose a novel, multistage fusion method where the two information streams are integrated in several layers of a deep neural network (DNN), and contrast it with a single-stage one where the streams are merged in a single point. Both methods depend on extracting summary linguistic embeddings from a pre-trained BERT model, and conditioning one or more intermediate representations of a convolutional model operating on log-Mel spectrograms. Experiments on the widely used IEMOCAP and MSP-Podcast databases demonstrate that the two fusion methods clearly outperform a shallow (late) fusion baseline and their unimodal constituents, both in terms of quantitative performance and qualitative behaviour. Our accompanying analysis further reveals a hitherto unexplored role of the underlying dialogue acts on unimodal and bimodal SER, with different models showing a biased behaviour across different acts. Overall, our multistage fusion shows better quantitative performance, surpassing all alternatives on most of our evaluations. This illustrates the potential of multistage fusion in better assimilating text and audio information.
As virtual assistants have become more diverse and specialized, so has the demand for application or brand-specific wake words. However, the wake-word-specific datasets typically used to train wake-word detectors are costly to create. In this paper, we explore two techniques to leverage acoustic modeling data for large-vocabulary speech recognition to improve a purpose-built wake-word detector: transfer learning and knowledge distillation. We also explore how these techniques interact with time-synchronous training targets to improve detection latency. Experiments are presented on the open-source "Hey Snips" dataset and a more challenging in-house far-field dataset. Using phone-synchronous targets and knowledge distillation from a large acoustic model, we are able to improve accuracy across dataset sizes for both datasets while reducing latency.
Speech directed to children differs from adult-directed speech in linguistic aspects such as repetition, word choice, and sentence length, as well as in aspects of the speech signal itself, such as prosodic and phonemic variation. Human language acquisition research indicates that child-directed speech helps language learners. This study explores the effect of child-directed speech when learning to extract semantic information from speech directly. We compare the task performance of models trained on adult-directed speech (ADS) and child-directed speech (CDS). We find indications that CDS helps in the initial stages of learning, but eventually, models trained on ADS reach comparable task performance, and generalize better. The results suggest that this is at least partially due to linguistic rather than acoustic properties of the two registers, as we see the same pattern when looking at models trained on acoustically comparable synthetic speech.
It has been established that Speech Affect Recognition for low resource languages is a difficult task. Here we present a Transfer learning based Speech Affect Recognition approach in which: we pre-train a model for high resource language affect recognition task and fine tune the parameters for low resource language using Deep Residual Network. Here we use standard four data sets to demonstrate that transfer learning can solve the problem of data scarcity for Affect Recognition task. We demonstrate that our approach is efficient by achieving 74.7 percent UAR on RAVDESS as source and Urdu data set as a target. Through an ablation study, we have identified that pre-trained model adds most of the features information, improvement in results and solves less data issues. Using this knowledge, we have also experimented on SAVEE and EMO-DB data set by setting Urdu as target language where only 400 utterances of data is available. This approach achieves high Unweighted Average Recall (UAR) when compared with existing algorithms.
We launched a community platform for collecting the ATC speech world-wide in the ATCO2 project. Filtering out unseen non-English speech is one of the main components in the data processing pipeline. The proposed English Language Detection (ELD) system is based on the embeddings from Bayesian subspace multinomial model. It is trained on the word confusion network from an ASR system. It is robust, easy to train, and light weighted. We achieved 0.0439 equal-error-rate (EER), a 50% relative reduction as compared to the state-of-the-art acoustic ELD system based on x-vectors, in the in-domain scenario. Further, we achieved an EER of 0.1352, a 33% relative reduction as compared to the acoustic ELD, in the unseen language (out-of-domain) condition. We plan to publish the evaluation dataset from the ATCO2 project.
This paper focuses on finding suitable features to robustly recognize emotions and evaluate customer satisfaction from speech in real acoustic scenarios. The classification of emotions is based on standard and well-known corpora and the evaluation of customer satisfaction is based on recordings of real opinions given by customers about the received service during phone calls with call-center agents. The feature sets considered in this study include two speaker models, namely x-vectors and i-vectors, and also the well known feature set introduced in the Interspeech 2010 Paralinguistics Challenge (I2010PC). Additionally, we introduce the use of phonation, articulation and prosody features extracted with the DisVoice framework as alternative feature sets to robustly model emotions and customer satisfaction from speech. The results indicate that the I2010PC feature set is the best approach to classify emotions in the standard databases typically used in the literature. When considering the recordings collected in the call-center, without any control over the acoustic conditions, the best results are obtained with our articulation features. The I2010PC feature set includes 1584 measures while the articulation approach only includes 488 measures. We think that the proposed approach is more suitable for real-world applications where the acoustic conditions are not controlled and also it is potentially more convenient for industrial applications.
In spoken language understanding (SLU), what the user says is converted to his/her intent. Recent work on end-to-end SLU has shown that accuracy can be improved via pre-training approaches. We revisit ideas presented by Lugosch et al. using speech pre-training and three-module modeling; however, to ease construction of the end-to-end SLU model, we use as our phoneme module an open-source acoustic-phonetic model from a DNN-HMM hybrid automatic speech recognition (ASR) system instead of training one from scratch. Hence we fine-tune on speech only for the word module, and we apply multi-target learning (MTL) on the word and intent modules to jointly optimize SLU performance. MTL yields a relative reduction of 40% in intent-classification error rates (from 1.0% to 0.6%). Note that our three-module model is a streaming method. The final outcome of the proposed three-module modeling approach yields an intent accuracy of 99.4% on FluentSpeech, an intent error rate reduction of 50% compared to that of Lugosch et al. Although we focus on real-time streaming methods, we also list non-streaming methods for comparison.
Silent speech interfaces (SSI) aim to reconstruct the speech signal from a recording of the articulatory movement, such as an ultrasound video of the tongue. Currently, deep neural networks are the most successful technology for this task. The efficient solution requires methods that do not simply process single images, but are able to extract the tongue movement information from a sequence of video frames. One option for this is to apply recurrent neural structures such as the long short-term memory network (LSTM) in combination with 2D convolutional neural networks (CNNs). Here, we experiment with another approach that extends the CNN to perform 3D convolution, where the extra dimension corresponds to time. In particular, we apply the spatial and temporal convolutions in a decomposed form, which proved very successful recently in video action recognition. We find experimentally that our 3D network outperforms the CNN+LSTM model, indicating that 3D CNNs may be a feasible alternative to CNN+LSTM networks in SSI systems.
We present a preprocessed, ready-to-use automatic speech recognition corpus, BembaSpeech, consisting over 24 hours of read speech in the Bemba language, a written but low-resourced language spoken by over 30% of the population in Zambia. To assess its usefulness for training and testing ASR systems for Bemba, we train an end-to-end Bemba ASR system by fine-tuning a pre-trained DeepSpeech English model on the training portion of the BembaSpeech corpus. Our best model achieves a word error rate (WER) of 54.78%. The results show that the corpus can be used for building ASR systems for Bemba. The corpus and models are publicly released at https://github.com/csikasote/BembaSpeech.
End-to-end automatic speech recognition systems have achieved great accuracy by using deeper and deeper models. However, the increased depth comes with a larger receptive field that can negatively impact model performance in streaming scenarios. We propose an alternative approach that we call Neural Mixture Model. The basic idea is to introduce a parallel mixture of shallow networks instead of a very deep network. To validate this idea we design CarneliNet -- a CTC-based neural network composed of three mega-blocks. Each mega-block consists of multiple parallel shallow sub-networks based on 1D depthwise-separable convolutions. We evaluate the model on LibriSpeech, MLS and AISHELL-2 datasets and achieved close to state-of-the-art results for CTC-based models. Finally, we demonstrate that one can dynamically reconfigure the number of parallel sub-networks to accommodate the computational requirements without retraining.