Self-supervised pretrained models exhibit competitive performance in automatic speech recognition (ASR) on finetuning, even with limited in-domain supervised data. However, popular pretrained models are not suitable for streaming ASR because they are trained with full attention context. In this paper, we introduce XLSR-Transducer, where the XLSR-53 model is used as encoder in transducer setup. Our experiments on the AMI dataset reveal that the XLSR-Transducer achieves 4% absolute WER improvement over Whisper large-v2 and 8% over a Zipformer transducer model trained from scratch. To enable streaming capabilities, we investigate different attention masking patterns in the self-attention computation of transformer layers within the XLSR-53 model. We validate XLSR-Transducer on AMI and 5 languages from CommonVoice under low-resource scenarios. Finally, with the introduction of attention sinks, we reduce the left context by half while achieving a relative 12% improvement in WER.

In real-world speech data processing, the scarcity of annotated data and the abundance of unlabelled speech data present a significant challenge. To address this, we propose an efficient data selection pipeline for fine-tuning ASR models by generating pseudo-labels using WhisperX pipeline and selecting efficient labels for fine-tuning. In our work, we propose a domain classifier system developed with a computationally inexpensive TFIDF and classical machine learning algorithm. Later, we filter data from the classifier output using a novel metric that assesses word ratio and perplexity distribution. The filtered pseudo labels are then used for fine-tuning standard encoder-decoder Whisper models and Zipformer. Our proposed data selection pipeline reduces the dataset size by approximately 1/100thwhile maintaining performance comparable to the full dataset, outperforming random domain-independent selection strategies.

Shashi Kumar, Srikanth Madikeri, Juan Pablo Zuluaga Gomez, Iuliia Thorbecke, Esaú Villatoro-tello, Sergio Burdisso, Petr Motlicek, Karthik Pandia D S, Aravind Ganapathiraju. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Traditionally, automatic speech recognition (ASR) and speaker change detection (SCD) systems have been independently trained to generate comprehensive transcripts accompanied by speaker turns. Recently, joint training of ASR and SCD systems, by inserting speaker turn tokens in the ASR training text, has been shown to be successful. In this work, we present a multitask alternative to the joint training approach. Results obtained on the mix-headset audios of AMI corpus show that the proposed multitask training yields an absolute improvement of 1.8% in coverage and purity based F1 score on SCD task without ASR degradation. We also examine the trade-offs between the ASR and SCD performance when trained using multitask criteria. Additionally, we validate the speaker change information in the embedding spaces obtained after different transformer layers of a self-supervised pre-trained model, such as XLSR-53, by integrating an SCD classifier at the output of specific transformer layers. Results reveal that the use of different embedding spaces from XLSR-53 model for multitask ASR and SCD is advantageous.1

Spoken Language Understanding (SLU) technologies have greatly improved due to the effective pretraining of speech representations. A common requirement of industry-based solutions is the portability to deploy SLU models in voice-assistant devices. Thus, distilling knowledge from large text-based language models has become an attractive solution for achieving good performance and guaranteeing portability. In this paper, we introduce a novel architecture that uses a cross-modal attention mechanism to extract bin-level contextual embeddings from a word-confusion network (WNC) encoding such that these can be directly compared and aligned with traditional text-based contextual embeddings. This alignment is achieved using a recently proposed tokenwise constrastive loss function. We validate our architecture’s effectiveness by fine-tuning our WCN-based pretrained model to do intent classification (IC) on the well-known SLURP dataset. Obtained accuracy on the IC task (81%), depicts a 9.4% relative improvement compared to a recent/equivalent E2E method.

In this paper, we perform an exhaustive evaluation of different representations to address the intent classification problem in a Spoken Language Understanding (SLU) setup. We benchmark three types of systems to perform the SLU intent detection task: 1) text-based, 2) lattice-based, and a novel 3) multimodal approach. Our work provides a comprehensive analysis of what could be the achievable performance of different state-of-the-art SLU systems under different circumstances, e.g., automatically- vs. manually-generated transcripts. We evaluate the systems on the publicly available SLURP spoken language resource corpus. Our results indicate that using richer forms of Automatic Speech Recognition (ASR) outputs, namely word-consensus-networks, allows the SLU system to improve in comparison to the 1-best setup (5.5% relative improvement). However, crossmodal approaches, i.e., learning from acoustic and text embeddings, obtains performance similar to the oracle setup, a relative improvement of 17.8% over the 1-best configuration, being a recommended alternative to overcome the limitations of working with automatically generated transcripts.

In this paper, we evaluate different alternatives to process richer forms of Automatic Speech Recognition (ASR) output based on lattice expansion algorithms for Spoken Document Retrieval (SDR). Typically, SDR systems employ ASR transcripts to index and retrieve relevant documents. However, ASR errors negatively affect the retrieval performance. Multiple alternative hypotheses can also be used to augment the input to document retrieval to compensate for the erroneous one-best hypothesis. In Weighted Finite State Transducer-based ASR systems, using the n-best output (i.e. the top "n'' scoring hypotheses) for the retrieval task is common, since they can easily be fed to a traditional Information Retrieval (IR) pipeline. However, the n-best hypotheses are terribly redundant, and do not sufficiently encapsulate the richness of the ASR output, which is represented as an acyclic directed graph called the lattice. In particular, we utilize the lattice's constrained minimum path cover to generate a minimum set of hypotheses that serve as input to the reranking phase of IR. The novelty of our proposed approach is the incorporation of the lattice as an input for neural reranking by considering a set of hypotheses that represents every arc in the lattice. The obtained hypotheses are encoded through sentence embeddings using BERT-based models, namely SBERT and RoBERTa, and the final ranking of the retrieved segments is obtained with a max-pooling operation over the computed scores among the input query and the hypotheses set. We present our evaluation on the publicly available AMI meeting corpus. Our results indicate that the proposed use of hypotheses from the expanded lattice improves the SDR performance significantly over the n-best ASR output.

Most large vocabulary continuous speech recognition (LVCSR) systems in the past decade have used a context-dependent (CD) phone as the fundamental acoustic unit. We present one of the first robust LVCSR systems that uses a syllable-level acoustic unit for LVCSR on telephone-bandwidth speech. This effort is motivated by the inherent limitations in phone-based approaches-namely the lack of an easy and efficient way for modeling long-term temporal dependencies. A syllable unit spans a longer time frame, typically three phones, thereby offering a more parsimonious framework for modeling pronunciation variation in spontaneous speech. We present encouraging results which show that a syllable-based system exceeds the performance of a comparable triphone system both in terms of word error rate (WER) and complexity. The WER of the best syllabic system reported here is 49.1% on a standard Switchboard evaluation, a small improvement over the triphone system. We also report results on a much smaller recognition task, OGI Alphadigits, which was used to validate some of the benefits syllables offer over triphones. The syllable-based system exceeds the performance of the triphone system by nearly 20%, an impressive accomplishment since the alphadigits application consists mostly of phone-level minimal pair distinctions.

In this paper, we describe the use of a powerful machine learning scheme, Support Vector Machines (SVM), within the framework of hidden Markov model (HMM) based speech recognition. The hybrid SVM/HMM system has been developed based on our public domain toolkit. The hybrid system has been evaluated on the OGI Alphadigits corpus and performs at 11.6% WER, as compared to 12.7% with a triphone mixture-Gaussian HMM system, while using only a fifth of the training data used by triphone system. Several important issues that arise out of the nature of SVM classifiers have been addressed. We are in the process of migrating this technology to large vocabulary recognition tasks like SWITCHBOARD. 1. INTRODUCTION Speech recogn i t i on can be v i ewed as a pa t t ern recognition problem where we desire each unique sound t o be d i s t i ngu i shab l e f r om a l l o t he r sounds . Traditionally statistical models, such as Gaussian mixture models, have been used to "represent" th...

The lack of freely available state-of-the-art Speech-to-Text (STT) software has been a major hindrance to the development of new audio information processing technology. The high cost of the infrastructure required to conduct state-of-the-art speech recognition research prevents many small research groups from evaluating new ideas on large-scale tasks. In this paper, we present the core components of an available state-of-the-art STT system: an acoustic processor which converts the speech signal into a sequence of feature vectors; a training module which estimates the parameters for a Hidden Markov Model; a linguistic processor which predicts the next word given a sequence of previously recognized words; and a search engine which finds the most probable word sequence given a set of feature vectors. 1.

We present a set of experiments which explore the use of syllables for recognition of continuous alphadigit utterances. In this system, syllables are used as the primary unit of recognition. This work was motivated by our need to verify and isolate phenomena seen when performing syllable-based experiments on the Switchboard corpus. The performance of our base syllable system is better than a crossword triphone system while requiring a small portion of the resources necessary for triphone systems. All experiments were performed on the OGI Alphadigits corpus, which consists of telephone-bandwidth alphadigit strings. The word error rate (WER) of the best syllable system (context-independent syllables) reported here is 11.1% compared to 12.2% for a crossword triphone system.

The SWITCHBOARD (SWB) corpus is one of the most important benchmarks for recognition tasks involving large vocabulary conversational speech (LVCSR). The high error rates on SWB are largely attributable to an acoustic model mismatch, the high frequency of poorly articulated monosyllabic words, and large variations in pronunciations. It is imperative to improve the quality of segmentations and transcriptions of the training data to achieve better acoustic modeling. By adapting existing acoustic models to only a small subset of such improved transcriptions, we have achieved a 2% absolute improvement in performance.

Hidden Markov models (HMM) with Gaussian mixture observation densities are the dominant approach in speech recognition. These systems typically use a representational model for acoustic modeling which can often be prone to overfitting and does not translate to improved discrimination. We propose a new paradigm centered on principles of structural risk minimization using a discriminative framework for speech recognition based on support vector machines (SVMs). SVMs have the ability to simultaneously optimize the representational and discriminative ability of the acoustic classifiers. We have developed the first SVM-based large vocabulary speech recognition system that improves performance over traditional HMM-based systems. This hybrid system achieves a state-of-the-art word error rate of 10.6% on a continuous alphadigit task—a 10% improvement relative to an HMM system. On SWITCHBOARD, a large vocabulary task, the system improves performance over a traditional HMM system from 41.6% word error rate to 40.6%. This dissertation discusses several practical issues that arise when SVMs are incorporated into the hybrid system.

The p r imary p rob l em in l a rge vocabu l a ry conversational speech recognition (LVCSR) is poor acoustic-level matching due to large variability in pronunciations. There is much to explore about the “quality” of states in an HMM and the interrelationships between inter-state and intra-state Gaussians used to model speech. Of particular interest is the variable discriminating power of the individual states. The fundamental concept addressed in this paper is to investigate means of exploiting such dependencies through model topology optimization based on the Bayesian Information Criterion (BIC) and the Minimum Description Length (MDL) principle.

By building acoustic phonetic models which explicitly represent as much knowledge of pronunciation in a small domain (the digits) as possible, we can create a recognition system which not only performs well but allows for meaningful error analysis and improvement. An HMM-based recognizer for the digits and a few associated words was constructed in accord with these principles. About 65 phonetic models were trained on 140 carefully labeled utterances, then iteratively trained on unlabeled data under orthographic supervision. The basic system achieved less than 3% word error rate on digit strings of unknown length from unseen test speakers, and 1.4% on 7-digit strings of known length. This is competitive with word-based models using the same HMM engine and similar parameter settings. As an R&D system, it allows meaningful analysis of errors and relatively straightforward means of improvement.