Abstract

Due to the lack of a large annotated speech corpus, many low-resource Indian languages struggle to utilize recent advancements in deep neural network architectures for Automatic Speech Recognition (ASR) tasks. Collecting large-scale databases is an expensive and time-consuming task. Current approaches lack extensive traditional expert-based data acquisition guidelines, as they are tedious and complex. In this work, we present the International Institute of Information Technology Hyderabad-Crowd Sourced Telugu Database (IIITH-CSTD), a Telugu corpus collected through crowdsourcing

Abstract

This work proposes the utilization of a self-supervised pretrained network for developing a Language Identification (LID) system catering to low-resource Indian languages. The framework employed is Wav2vec2.0-XLSR-53, pre-trained on 53k hours of unlabeled speech data. The unsupervised training of the model enables it to learn the acoustic patterns specific to a language. Given that languages share phonetic space, multi-lingual pre-training is instrumental in learning crosslingual information and building systems that cater to low-resource languages. The results showcase a relative improvement of 33.2% over the DNN-A (DNN with attention) model and 19.04% over Dense Resnets for the Language Identification task on the IIITH-ILSC database using the proposed features.

Abstract

Stuttering is a prevalent speech disorder that affects millions of people worldwide. In this Show and Tell presentation, we demonstrate a novel platform that takes speech samples in English and Kannada to detect and analyze stuttering in patients. The user-friendly interface includes demographic details and speech samples, generating comprehensive reports for different stuttering disfluencies. The platform has four different user types, providing full read-only access for admins and full write access for super admins. Our platform provides valuable assistance for speechlanguage pathologists to evaluate speech samples. The proposed platform supports both live and recorded speech samples and presents a flexible approach to stuttering detection and analysis. Our research demonstrates the potential of technology to improve speech-language pathology for stuttering. Used F-score as a metric for evaluating the models for the stutter detection task.

Abstract

Hardware accelerators are being designed to offload compute-intensive tasks such as deep neural networks from the CPU to improve the overall performance of an application, specifically on the performance-per-watt metric. Encoder-decoderbased sequence-to-sequence models such as the Transformer model have demonstrated state-of-the-art results in end-to-end automatic speech recognition systems (ASRs). The Transformer model being intensive on memory and computation poses a challenge for an FPGA implementation. This paper proposes an end-to-end architecture to accelerate a Transformer for an ASR system. The host CPU orchestrates the computations from different encoder and decoder stages of the Transformer architecture on the designed hardware accelerator with no necessity for intervening FPGA reconfiguration. The communication latency is hidden by prefetching the weights of the next encoder/decoder block while the current block is being processed. The computation is split across both the Super Logic Regions (SLRs) of the FPGA, mitigating the inter-SLR communication. The proposed design presents an optimal latency, exploiting the available resources. The accelerator design is realized using highlevel synthesis tools and evaluated on an Alveo U-50 FPGA card. The design demonstrates an average speed-up of 32× compared to an Intel Xeon E5-2640 CPU and 8.8× compared to NVIDIA GeForce RTX 3080 Ti Graphics card for a 32-bit floating point single precision model. Index Terms—Automatic speech recognition, Hardware accelerator, Transformer, FPGAs.

Abstract

Multilingual Automatic Speech Recognition (ASR) for Indian lan- guages is an obvious technique for leveraging their similarities. We present a detailed analysis of how phonological similarities and differences between languages affect Time Delay Neural Net- work (TDNN) and End-to-End (E2E) ASR. To study this, we select genealogically similar pairs from five Indian languages and train bilingual acoustic models. We compare these against corresponding monolingual acoustic models and find similar phoneme distribu- tions within speech to be the primary factor for improving model performance, with phoneme overlap being secondary. The influ- ence of phonological properties on performance is visible in both cases. Word Error Rate (WER) of E2E decreased by a median of 2.35%, and upto 8.5% when the phonological similarity was great- est. WER of TDNN increased by 11.69% when the similarity was lowest. Thus, it is clear that the choice of supplementary language is important for model performance. Index Terms— Phonetics and Phonology, Analysis, Automatic Speech Recognition, Low-Resource Languages, Multilingual Sys-

Abstract

Conventional Automatic Speech Recognition (ASR) systems are susceptible to dialect variations within a language, thereby adversely affecting the ASR. Therefore, the current practice is to use dialect-specific ASRs. However, dialectspecific information or data is hard to obtain making it difficult to build dialect-specific ASRs. Furthermore, it is cumbersome to maintain multiple dialect-specific ASR systems for each language. We build a unified multi-dialect End-to-End ASR that removes the need for a dialect recognition block and the need to maintain multiple dialect-specific ASRs for three Telugu regional dialects: Telangana, Coastal Andhra, and Rayalaseema. We find that pooling the data and training a multi-dialect ASR benefits the low-resource dialect the most – an improvement of over 9.71% in relative Word Error Rate (WER). Subsequently, we experiment with multi-task ASRs where the primary task is to transcribe the audio and the secondary task is to predict the dialect. We do this by adding a Dialect ID to the output targets. Such a model outperforms naive multi-dialect ASRs by up to 8.24% in relative WER. Additionally, we test this model on a dialect recognition task and find that it outperforms strong baselines by 6.14% in accuracy. Index Terms: multi-dialect, speech recognition, sequence-tosequence, dialect recognition

Abstract

Parkinson’s disease (PD) is a speech disorder caused by the effect of neurological damage on motor control activity of speech production. Due to the noninvasive nature, detection of PD from the speech is remarkable. This study proposes the use of Long-Term Average Spectral (LTAS) feature to detect PD from speech using different filterbank techniques and classifiers on the PC-GITA dataset. The fact that PD can be perceived from long-temporal information has motivated us to explore the LTAS features for PD detection. Performance of LTAS features is compared to baseline features, namely statistical and openSMILE features (ComParE, GeMAPS, and eGeMAPS). The performance of these features is validated using five different classifiers. The LTAS features, which are obtained with auditory filterbanks (Constant Q Transform and Gammatone), showed better performance as compared to the statistical features (MFCC_Stat and LPCC_Stat) and openSMILE features (ComParE, GeMAPS, and eGeMAPS). The combination of the Gammatone-based LTAS feature with eGeMAPS further improved the detection accuracy of the PD detection system on the PC-GITA database.

Abstract

Previous research has found that Acoustic Models (AM) of an Automatic Speech Recognition (ASR) system are susceptible to dialect variations within a language, thereby adversely affecting the ASR. To counter this, researchers have proposed to build a dialectspecific AM while keeping the Language Model (LM) constant for all the dialects. This study explores the effect of dialect mismatched LM by considering three different Telugu regional dialects: Telangana, Coastal Andhra, and Rayalaseema. We show that dialect variations that surface in the form of a different lexicon, grammar, and occasionally semantics can significantly degrade the performance of the LM under mismatched conditions. Therefore, this degradation has an adverse effect on the ASR even when dialect-specific AM is used. We show a degradation of up to 13.13 perplexity points when LM is used under mismatched conditions. Furthermore, we show a degradation of over 9% and over 15% in Character Error Rate (CER) and Word Error Rate (WER), respectively, in the ASR systems when using mismatched LMs over matched LMs

Abstract

Alzheimer’s Dementia is a progressive neurological disorder characterized by cognitive impairment. It affects mem- ory, thinking skills, language, and the ability to perform simple tasks. Detection of Alzheimer’s Dementia from the speech is considered a primitive task, as most speech cues are preserved in it. Studies in the literature focused mainly on the lexical features and few acoustic features for detecting Alzheimer’s disease. The present work explores the single frequency filtering cepstral coefficients (SFCC) for the automatic detection of Alzheimer’s disease. In contrast to STFTs, the proposed feature has better temporal and spectral resolution and captures the transient part more appropriately. This offers a very compact and efficient way to derive the formant structure in the speech signal. The experiments were conducted on the ADReSSo dataset, using the support vector machine classifier. The classification performance was compared with several baseline features like Mel-frequency cepstral coefficients (MFCC), perceptual linear prediction (PLP), linear prediction cepstral coefficient (LPCC), Mel frequency cepstral coefficients of LP-residual (MFCC-WR), ZFF signal (MFCC-ZF) and eGeMAPS (openSMILE). The experiments conducted on Alzheimer’s Dementia classification task show that the proposed feature performs better than conventional MFCCs. Among all the features, SFCC offers the best classification accuracy of 65.1% and 60.6% for dementia detection on cross- validation and test data, respectively. The combination of baseline features with SFCC features further improved the performance. Index Terms—Alzheimer’s disease, Cognitive impairment, Sin- gle frequency filtering cepstral coefficients.

Abstract

Parkinson’s disease (PD) is a progressive neurodegenerative dis- order of the central nervous system identified by motor and non- motor activities abnormalities. PD affects respiration, laryngeal, articulation, resonance, and prosodic aspects of speech production. Detection of PD from speech is a non-invasive approach useful for automatic screening. Perceptual attributes of speech due to PD are manifested as temporal variations in speech. In this regard, current work investigated the use of LSTM and BiLSTM networks with shifted delta cepstral (SDC) features to detect PD from speech. Further in BiLSTM networks, a multi-head attention mechanism is introduced, assuming that each head captures distinct information to detect PD. SDC features obtained from MFCCs, and SFFCCs are used for developing the PD detection system. The performance of the experiments is validated using the PC-GITA database. The experimental results revealed that BiLSTM networks give a rela- tive improvement of 4-5% over the LSTM networks. The use of a multi-head attention mechanism further improved the detection accuracy of the PD detection system, showing that it can capture various discriminative features.