We introduce RESPIN-S1.0, the largest publicly available dialect-rich read-speech corpus for Indian languages, comprising more than 10,000 hours of validated audio across nine major languages: Bengali, Bhojpuri, Chhattisgarhi, Hindi, Kannada, Magahi, Maithili, Marathi, and Telugu. Indian languages exhibit high dialectal variation and are spoken by populations that remain digitally underserved. Existing speech corpora typically represent only standard dialects and lack domain and linguistic diversity. RESPIN-S1.0 addresses this limitation by collecting speech across more than 38 dialects and two high-impact domains: agriculture and finance. Text data were composed by native dialect speakers and validated through a pipeline combining automated and manual checks. Over 200,000 unique sentences were recorded through a crowdsourced mobile platform and categorised into clean, semi-noisy, and noisy subsets based on transcription quality, with the clean portion alone exceeding 10,000 hours. Along with audio and transcriptions, RESPIN provides dialect-aware phonetic lexicons, speaker metadata, and reproducible train, development, and test splits. To benchmark performance, we evaluate multiple ASR models, including TDNN-HMM, E-Branchformer, Whisper, and wav2vec2-based self-supervised models, and find that fine-tuning on RESPIN significantly improves recognition accuracy over pretrained baselines. A subset of RESPIN-S1.0 has already supported community challenges such as the SLT Code Hackathon 2022 and MADASR@ASRU 2023 and 2025, releasing more than 1,200 hours publicly. This resource supports research in dialectal ASR, language identification, and related speech technologies, establishing a comprehensive benchmark for inclusive, dialect-rich ASR in multilingual low-resource settings. Dataset: https://spiredatasets.ee.iisc.ac.in/respincorpus Code: https://github.com/labspire/respin_baselines.git

In this work, several methods have been proposed towards improving the performance of dialectal automatic speech recognition (ASR). A novel encoder architecture has been introduced that is suited for multi-dialect ASR training. Further, we propose Multi-Task Self-Supervised learning (SSL) fine-tuning using CTC and dialect identification. Additionally, the use of different language models (LM) to improve the performance of dialectal ASR has been investigated. Around 800 hours of Bengali and Bhojpuri data, released as a part of the MADASR ASRU challenge have been used to train these models. The work shows that the proposed multi-encoder ASR observes a relative reduction of 7.5% and 9% in WER in Bhojpuri and Bengali, respectively. Additionally, we also observe a 1-2% WER reduction in fine-tuning SSL, further improving performance in these languages. Moreover, we observe advantages in using dialect-specific LM decoding based on predicted dialect.

Recently, there is increasing interest in multilingual automatic speech recognition (ASR) where a speech recognition system caters to multiple low resource languages by taking advantage of low amounts of labeled corpora in multiple languages. With multilingualism becoming common in today's world, there has been increasing interest in code-switching ASR as well. In code-switching, multiple languages are freely interchanged within a single sentence or between sentences. The success of low-resource multilingual and code-switching ASR often depends on the variety of languages in terms of their acoustics, linguistic characteristics as well as the amount of data available and how these are carefully considered in building the ASR system. In this challenge, we would like to focus on building multilingual and code-switching ASR systems through two different subtasks related to a total of seven Indian languages, namely Hindi, Marathi, Odia, Tamil, Telugu, Gujarati and Bengali. For this purpose, we provide a total of ~600 hours of transcribed speech data, comprising train and test sets, in these languages including two code-switched language pairs, Hindi-English and Bengali-English. We also provide a baseline recipe for both the tasks with a WER of 30.73% and 32.45% on the test sets of multilingual and code-switching subtasks, respectively.

In this study, we evaluate and compare two different approaches for multilingual phone recognition in code-switched and non-code-switched scenarios. First approach is a front-end Language Identification (LID)-switched to a monolingual phone recognizer (LID-Mono), trained individually on each of the languages present in multilingual dataset. In the second approach, a common multilingual phone-set derived from the International Phonetic Alphabet (IPA) transcription of the multilingual dataset is used to develop a Multilingual Phone Recognition System (Multi-PRS). The bilingual code-switching experiments are conducted using Kannada and Urdu languages. In the first approach, LID is performed using the state-of-the-art i-vectors. Both monolingual and multilingual phone recognition systems are trained using Deep Neural Networks. The performance of LID-Mono and Multi-PRS approaches are compared and analysed in detail. It is found that the performance of Multi-PRS approach is superior compared to more conventional LID-Mono approach in both code-switched and non-code-switched scenarios. For code-switched speech, the effect of length of segments (that are used to perform LID) on the performance of LID-Mono system is studied by varying the window size from 500 ms to 5.0 s, and full utterance. The LID-Mono approach heavily depends on the accuracy of the LID system and the LID errors cannot be recovered. But, the Multi-PRS system by virtue of not having to do a front-end LID switching and designed based on the common multilingual phone-set derived from several languages, is not constrained by the accuracy of the LID system, and hence performs effectively on code-switched and non-code-switched speech, offering low Phone Error Rates than the LID-Mono system.

We address the problem of efficient acoustic-model refinement (continuous retraining) using semi-supervised and active learning for a low resource Indian language, wherein the low resource constraints are having i) a small labeled corpus from which to train a baseline `seed' acoustic model and ii) a large training corpus without orthographic labeling or from which to perform a data selection for manual labeling at low costs. The proposed semi-supervised learning decodes the unlabeled large training corpus using the seed model and through various protocols, selects the decoded utterances with high reliability using confidence levels (that correlate to the WER of the decoded utterances) and iterative bootstrapping. The proposed active learning protocol uses confidence level based metric to select the decoded utterances from the large unlabeled corpus for further labeling. The semi-supervised learning protocols can offer a WER reduction, from a poorly trained seed model, by as much as 50% of the best WER-reduction realizable from the seed model's WER, if the large corpus were labeled and used for acoustic-model training. The active learning protocols allow that only 60% of the entire training corpus be manually labeled, to reach the same performance as the entire data.

We consider the problem of automatically detecting if a speaker is suffering from common cold from his/her speech. When a speaker has symptoms of cold, his/her voice quality changes compared to the normal one. We hypothesize that such a change in voice quality could be reflected in lower likelihoods from a model built using normal speech. In order to capture this, we compute a 120-dimensional posteriorgram feature in each frame using Gaussian mixture model from 120 states of 40 three-states phonetic hidden Markov models trained on approximately 16.4 hours of normal English speech. Finally, a fixed 5160-dimensional phoneme state posteriorgram (PSP) feature vector for each utterance is obtained by computing statistics from the posteriorgram feature trajectory. Experiments on the 2017-Cold sub-challenge data show that when the decisions from bag-of-Audio-words (BoAW) and end-To-end (e2e) are combined with those from PSP features with unweighted majority rule, the UAR on the development set becomes 69 which is 2.9 (absolute) better than the best of the UARs obtained by the baseline schemes. When the decisions from ComParE, BoAW and PSP features are combined with simple majority rule, it results in a UAR of 68.52 on the test set.