Features such as punctuation, capitalization, and formatting of entities are important for readability, understanding, and natural language processing tasks. However, Automatic Speech Recognition (ASR) systems produce spoken-form text devoid of formatting, and tagging approaches to formatting address just one or two features at a time. In this paper, we unify spoken-to-written text conversion via a two-stage process: First, we use a single transformer tagging model to jointly produce token-level tags for inverse text normalization (ITN), punctuation, capitalization, and disfluencies. Then, we apply the tags to generate written-form text and use weighted finite state transducer (WFST) grammars to format tagged ITN entity spans. Despite joining four models into one, our unified tagging approach matches or outperforms task-specific models across all four tasks on benchmark test sets across several domains.
Wake-up words (WUW) is a short sentence used to activate a speech recognition system to receive the user's speech input. WUW utterances include not only the lexical information for waking up the system but also non-lexical information such as speaker identity or emotion. In particular, recognizing the user's emotional state may elaborate the voice communication. However, there is few dataset where the emotional state of the WUW utterances is labeled. In this paper, we introduce Hi, KIA, a new WUW dataset which consists of 488 Korean accent emotional utterances collected from four male and four female speakers and each of utterances is labeled with four emotional states including anger, happy, sad, or neutral. We present the step-by-step procedure to build the dataset, covering scenario selection, post-processing, and human validation for label agreement. Also, we provide two classification models for WUW speech emotion recognition using the dataset. One is based on traditional hand-craft features and the other is a transfer-learning approach using a pre-trained neural network. These classification models could be used as benchmarks in further research.
Self-supervised learning (SSL) speech models such as wav2vec and HuBERT have demonstrated state-of-the-art performance on automatic speech recognition (ASR) and proved to be extremely useful in low label-resource settings. However, the success of SSL models has yet to transfer to utterance-level tasks such as speaker, emotion, and language recognition, which still require supervised fine-tuning of the SSL models to obtain good performance. We argue that the problem is caused by the lack of disentangled representations and an utterance-level learning objective for these tasks. Inspired by how HuBERT uses clustering to discover hidden acoustic units, we formulate a factor analysis (FA) model that uses the discovered hidden acoustic units to align the SSL features. The underlying utterance-level representations are disentangled from the content of speech using probabilistic inference on the aligned features. Furthermore, the variational lower bound derived from the FA model provides an utterance-level objective, allowing error gradients to be backpropagated to the Transformer layers to learn highly discriminative acoustic units. When used in conjunction with HuBERT's masked prediction training, our models outperform the current best model, WavLM, on all utterance-level non-semantic tasks on the SUPERB benchmark with only 20% of labeled data.
Though Dialogue State Tracking (DST) is a core component of spoken dialogue systems, recent work on this task mostly deals with chat corpora, disregarding the discrepancies between spoken and written language.In this paper, we propose OLISIA, a cascade system which integrates an Automatic Speech Recognition (ASR) model and a DST model. We introduce several adaptations in the ASR and DST modules to improve integration and robustness to spoken conversations.With these adaptations, our system ranked first in DSTC11 Track 3, a benchmark to evaluate spoken DST. We conduct an in-depth analysis of the results and find that normalizing the ASR outputs and adapting the DST inputs through data augmentation, along with increasing the pre-trained models size all play an important role in reducing the performance discrepancy between written and spoken conversations.
Self-supervised speech recognition models require considerable labeled training data for learning high-fidelity representations for Automatic Speech Recognition (ASR), which hinders their application to low-resource languages. We consider the task of identifying an optimal subset of training data to fine-tune self-supervised speech models for ASR. We make a surprising observation that active learning strategies for sampling harder-to-learn examples do not perform better than random subset selection for fine-tuning self-supervised ASR. We then present the COWERAGE algorithm for better subset selection in self-supervised ASR which is based on our finding that ensuring the coverage of examples based on training WER in the early training epochs leads to better generalization performance. Extensive experiments on the wav2vec 2.0 model and TIMIT dataset show the effectiveness of COWERAGE, with up to 27% absolute WER improvement over active learning methods. We also report the connection between training WER and the phonemic cover and demonstrate that our algorithm ensures inclusion of phonemically diverse examples.
Self-supervision has shown great potential for audio-visual speech recognition by vastly reducing the amount of labeled data required to build good systems. However, existing methods are either not entirely end-to-end or do not train joint representations of both modalities. In this paper, we introduce AV-data2vec which addresses these challenges and builds audio-visual representations based on predicting contextualized representations which has been successful in the uni-modal case. The model uses a shared transformer encoder for both audio and video and can combine both modalities to improve speech recognition. Results on LRS3 show that AV-data2vec consistently outperforms existing methods under most settings.
In this work, we present our deployment-ready Speech-to-Speech Machine Translation (SSMT) system for English-Hindi, English-Marathi, and Hindi-Marathi language pairs. We develop the SSMT system by cascading Automatic Speech Recognition (ASR), Disfluency Correction (DC), Machine Translation (MT), and Text-to-Speech Synthesis (TTS) models. We discuss the challenges faced during the research and development stage and the scalable deployment of the SSMT system as a publicly accessible web service. On the MT part of the pipeline too, we create a Text-to-Text Machine Translation (TTMT) service in all six translation directions involving English, Hindi, and Marathi. To mitigate data scarcity, we develop a LaBSE-based corpus filtering tool to select high-quality parallel sentences from a noisy pseudo-parallel corpus for training the TTMT system. All the data used for training the SSMT and TTMT systems and the best models are being made publicly available. Users of our system are (a) Govt. of India in the context of its new education policy (NEP), (b) tourists who criss-cross the multilingual landscape of India, (c) Indian Judiciary where a leading cause of the pendency of cases (to the order of 10 million as on date) is the translation of case papers, (d) farmers who need weather and price information and so on. We also share the feedback received from various stakeholders when our SSMT and TTMT systems were demonstrated in large public events.
Recent studies have shown that the benefits provided by self-supervised pre-training and self-training (pseudo-labeling) are complementary. Semi-supervised fine-tuning strategies under the pre-training framework, however, remain insufficiently studied. Besides, modern semi-supervised speech recognition algorithms either treat unlabeled data indiscriminately or filter out noisy samples with a confidence threshold. The dissimilarities among different unlabeled data are often ignored. In this paper, we propose Censer, a semi-supervised speech recognition algorithm based on self-supervised pre-training to maximize the utilization of unlabeled data. The pre-training stage of Censer adopts wav2vec2.0 and the fine-tuning stage employs an improved semi-supervised learning algorithm from slimIPL, which leverages unlabeled data progressively according to their pseudo labels' qualities. We also incorporate a temporal pseudo label pool and an exponential moving average to control the pseudo labels' update frequency and to avoid model divergence. Experimental results on Libri-Light and LibriSpeech datasets manifest our proposed method achieves better performance compared to existing approaches while being more unified.
Attention layers are an integral part of modern end-to-end automatic speech recognition systems, for instance as part of the Transformer or Conformer architecture. Attention is typically multi-headed, where each head has an independent set of learned parameters and operates on the same input feature sequence. The output of multi-headed attention is a fusion of the outputs from the individual heads. We empirically analyze the diversity between representations produced by the different attention heads and demonstrate that the heads become highly correlated during the course of training. We investigate a few approaches to increasing attention head diversity, including using different attention mechanisms for each head and auxiliary training loss functions to promote head diversity. We show that introducing diversity-promoting auxiliary loss functions during training is a more effective approach, and obtain WER improvements of up to 6% relative on the Librispeech corpus. Finally, we draw a connection between the diversity of attention heads and the similarity of the gradients of head parameters.
The emotion detection technology to enhance human decision-making is an important research issue for real-world applications, but real-life emotion datasets are relatively rare and small. The experiments conducted in this paper use the CEMO, which was collected in a French emergency call center. Two pre-trained models based on speech and text were fine-tuned for speech emotion recognition. Using pre-trained Transformer encoders mitigates our data's limited and sparse nature. This paper explores the different fusion strategies of these modality-specific models. In particular, fusions with and without cross-attention mechanisms were tested to gather the most relevant information from both the speech and text encoders. We show that multimodal fusion brings an absolute gain of 4-9% with respect to either single modality and that the Symmetric multi-headed cross-attention mechanism performed better than late classical fusion approaches. Our experiments also suggest that for the real-life CEMO corpus, the audio component encodes more emotive information than the textual one.