Kathbath: Indian Read-Speech Corpus

• Kathbath is a large-scale, human-verified read-speech corpus that provides 1,684 hours of clean speech across 12 Indian languages from a diverse pool of 1,218 speakers.
• It employs a robust collection protocol with controlled splits and human verification to ensure quality data for tasks like ASR, LID, SID, ASV, QbE, and KWS.
• Researchers use Kathbath in IndicSUPERB, Vistaar, and cross-corpus studies to analyze domain generalization, speaker variation, and acoustic robustness.

Searching arXiv for the cited Kathbath-related papers to ground the article in the primary sources. Kathbath is a large-scale, human-verified read-speech corpus for Indian languages and a central substrate for benchmark construction in Indic speech research. It contains 1,684 hours of labeled clean speech across 12 languages, contributed by 1,218 speakers from 203 districts spanning 22 Indian states, and serves as the core resource behind IndicSUPERB, which defines benchmarks for Automatic Speech Recognition, Speaker Verification, Speaker Identification, Language Identification, Query By Example, and Keyword Spotting. Subsequent work has also used Kathbath as one of the 59 benchmarks in Vistaar and, in a separate large-scale Indic LID study, strictly as an out-of-domain cross-corpus evaluation set relative to Vaani (Javed et al., 2022, Bhogale et al., 2023, Basu et al., 8 Jun 2026).

1. Corpus definition and linguistic coverage

Kathbath is a labeled read-speech dataset designed for Indian-language speech research. Its speech is clean, its utterances are aligned with transcripts, and its metadata includes language and speaker identity. The corpus covers 12 languages: Bengali (bn), Gujarati (gu), Hindi (hi), Kannada (kn), Malayalam (ml), Marathi (mr), Odia (or), Punjabi (pa), Sanskrit (sa), Tamil (ta), Telugu (te), and Urdu (ur). The dataset contains approximately 0.9 million unique utterances, with each sentence read by one speaker only and no repeats.

The per-language clean and noisy hours reported in IndicSUPERB are as follows.

Language	Clean hours	Noisy hours
Bengali (bn)	115.8	38.8
Gujarati (gu)	129.3	65.1
Hindi (hi)	150.2	47.7
Kannada (kn)	165.8	64.5
Malayalam (ml)	147.3	15.7
Marathi (mr)	185.2	86.5
Odia (or)	111.6	27.6
Punjabi (pa)	136.9	36.5
Sanskrit (sa)	115.5	76.3
Tamil (ta)	185.1	95.8
Telugu (te)	154.9	74.8
Urdu (ur)	86.7	77.0

The same source reports per-language speaker counts and vocabulary sizes, with overall gender representation described as approximately balanced and with deliberate over-representation of female speakers in some languages to study bias. A plausible implication is that Kathbath was designed not merely as an ASR corpus but as a controlled experimental resource for multi-task SLU under demographic and geographic variation (Javed et al., 2022).

2. Collection protocol and verification regime

Kathbath was collected as read speech because read-speech acquisition is cost-effective, scales well under remote collection, and permits tighter control over speaker metadata than curated web audio. Prompt sentences were sampled from IndicCorp, described as large monolingual corpora drawn from Indian news and government websites. Approximately 100,000 sentences per language were selected, restricted to alphanumeric characters and to lengths of 8–15 words to improve fluency and reduce pronunciation difficulty.

Collection used the Microsoft Karya Android platform in an offline, distributed setting during the COVID-19 pandemic. Access codes controlled 100-sentence batches per participant. Every recording then passed through a maker–checker human verification workflow. Verifiers assigned scores on three dimensions—accuracy, volume, and quality—each taking values from 0 to 2. Only recordings with a score of 2 on all three dimensions entered the clean set. The paper reports that approximately 60% of early data was rejected because of misreading and background noise; mitigations included asking speakers to practice reading, skip unclear sentences, and record at night to reduce ambient noise. Recordings were made in typical home environments, but the paper does not specify sampling rate, audio format, or microphone characteristics (Javed et al., 2022).

3. Splits, metadata, and evaluation conditions

Each utterance in Kathbath carries an aligned transcription, a language label, and a speaker ID; metadata also includes gender and district/state. The corpus was partitioned to support both clean/noisy and known-speaker/unknown-speaker evaluation conditions.

The clean split design is explicit. For each language, Test-Unknown holds out utterances from 10 male and 10 female speakers totaling approximately 3 hours, with all other recordings from those speakers removed from train and validation. Test-Known samples 10 male and 10 female speakers totaling approximately 5 hours, and these speakers may also appear in training. Validation follows the same structure as Test-Known. Training consists of the remaining clean recordings. No sentence repeats occur across splits by design.

Noisy partitions are derived from rejected recordings that still received scores of at least 1 on accuracy, volume, and quality. These include Noisy Test-Unknown, Noisy Test-Known, and Noisy Validation. This split design is important because it isolates speaker generalization and acoustic robustness rather than conflating them with lexical overlap. The same paper notes that balanced evaluation sets enable systematic bias analysis, including gender-skew effects on ASR performance (Javed et al., 2022).

4. Role in IndicSUPERB and task construction

IndicSUPERB uses Kathbath to instantiate six SLU tasks. ASR is built directly from aligned audio–text pairs. LID pools audio across the 12 languages and treats language ID as a 12-class target. SID appears in two forms: within-language speaker identification and multilingual speaker identification over the pooled speaker set. ASV is framed as binary verification using speaker embeddings and cosine similarity. QbE uses an additional read-speech collection centered on entity names, and KWS uses top-50 command/control keywords per language except Sanskrit.

Task	Construction from Kathbath	Metric
ASR	Aligned audio–text pairs; monolingual and multilingual fine-tuning	WER
LID	12-class classification over pooled languages	Accuracy
SID (mono/multi)	Speaker classification within language or across all languages	Accuracy
ASV	Same/different speaker verification from embeddings	EER
QbE	Read-speech retrieval over entity-name queries	MTWV
KWS	Keyword classification with SSL features and linear head	Accuracy

The metric definitions are explicit. For ASR, Word Error Rate is

$\mathrm{WER} = \frac{S + D + I}{N},$

where $S$ is substitutions, $D$ deletions, $I$ insertions, and $N$ the number of words in the reference. For LID, SID, and KWS, the metric is accuracy. For ASV, the metric is Equal Error Rate, defined at the operating point where False Acceptance Rate equals False Rejection Rate. For QbE, the metric is Maximum Term Weighted Value.

IndicSUPERB reports large gains from self-supervised features on Kathbath-derived tasks. For LID, averaged across languages, FBANK reaches 27% on Test-Clean Known and 14.10% on Test-Clean Unknown, whereas IndicWav2Vec reaches 98.24% and 90.78%, respectively; the paper highlights this as an approximately 76% improvement. For ASR with a 6-gram KenLM, monolingual fine-tuning yields average WERs of 12.4% on Test Known and 13.1% on Test Unknown for IndicWav2Vec, while multilingual fine-tuning yields 13.1% and 15.2%, respectively. The same benchmark also shows that performance degrades under unknown-speaker and noisy conditions across tasks, making Kathbath’s controlled split structure central to robustness analysis (Javed et al., 2022).

5. Kathbath in Vistaar and ASR benchmarking

In Vistaar, Kathbath is one of 59 benchmarks used to evaluate Indian-language ASR systems across diverse domains. There it is described as read speech collected on Android phones with the Karya application by AI4Bharat, with sentences sourced from Wikipedia and news articles. Kathbath spans all 12 Vistaar languages and is also included in Vistaar-Train, where the reported hours are bn 103, gu 116, hi 137, kn 153, ml 134, mr 172, or 99, pa 124, sa 102, ta 172, te 142, and ur 74, for a total of 1,527 hours.

Vistaar also defines Kathbath-Hard, produced by adding background noise from the ESC dataset with randomly chosen SNR between 3 dB and 30 dB. This provides a controlled robustness variant rather than a naturally noisy collection.

ASR results on Kathbath in Vistaar are reported in WER only. On the Hindi subset, the table of publicly available models gives: IndicWhisper 10.3, IndicWav2vec 12.2, Nvidia-large 12.7, Azure STT 13.6, Nvidia-medium 14.0, and Google STT 14.3. On Kathbath-Hard for Hindi, the corresponding WERs are 12.0, 16.2, 14.2, 15.1, 15.6, and 16.7. Cross-language Kathbath averages are also reported: Google STT 40.5, IndicWav2Vec 22.3, and IndicWhisper 22.2; on Kathbath-Hard, the averages become 42.4, 26.9, and 25.3. Azure STT is unavailable for some Kathbath languages and is marked NA for or, pa, sa, and ur. The paper reports WER only, with no CER, confidence intervals, or error-type decomposition (Bhogale et al., 2023).

6. Kathbath as an out-of-domain benchmark for Indic spoken LID

A later comparative study on large-scale Indic spoken language identification uses Kathbath in a different role. There, Kathbath is treated strictly as an out-of-domain, cross-corpus benchmark. All models are trained only on Vaani and then evaluated on Vaani-Test, FLEURS, and Kathbath; Kathbath contains 11 languages in this evaluation, although the paper does not enumerate them. The study motivates this setup by noting that Indic speech corpora differ in recording conditions, speaking style, and speaker demographics, making Kathbath a probe of domain generalization.

The model family is fixed: either Whisper (openai/whisper-medium, 350M parameters, 1024-dimensional frame-level representations) or FastConformer (ARTPARK-IISc/Vaani-FastConformer-Multilingual, 430M parameters, 1024-dimensional representations), followed by self-attention pooling and a single linear classification head over $L=42$ languages. Encoders are tested in frozen and fine-tuned modes. Training objectives are Cross-Entropy, CE + Supervised Contrastive loss, and Hierarchical Softmax over a four-level linguistic tree: Root $\rightarrow$ Family $\rightarrow$ Sub-family $\rightarrow$ Language, with families Indo-Aryan, Dravidian, Sino-Tibetan, and European. Optimization uses AdamW, with learning rates $10^{-5}$ for fine-tuning and $S$0 for linear probing, batch size 4 with gradient accumulation to effective batch size 96, and CE+SupCon hyperparameters $S$1 and $S$2.

The objective definitions are given as

$S$3

$S$4

$S$5

and for hierarchical softmax,

$S$6

$S$7

$S$8

Macro accuracy is

$S$9

Kathbath results are central to the paper’s domain-generalization argument. Under CE loss, frozen Whisper scores 57.7 macro accuracy and fine-tuned Whisper 68.3, whereas frozen FastConformer reaches 90.9 and fine-tuned FastConformer 87.4. For fine-tuned encoders across objectives, Whisper obtains 68.3 with CE, 71.0 with CE+SupCon, and 75.8 with HSM; FastConformer obtains 87.4 with CE, 79.8 with CE+SupCon, and 90.0 with HSM. External LID baselines on Kathbath are FBMMS at 91.3, supporting only 30/42 languages in Vaani, and SpeechBrain ECAPA-TDNN at 87.9, supporting 13/42. The paper emphasizes that frozen FastConformer achieves over 90% macro accuracy on FLEURS and Kathbath without any task-specific adaptation, that HSM consistently improves out-of-domain performance, and that CE+SupCon degrades FastConformer’s cross-corpus generalization by 7.6 points relative to CE on Kathbath. Its confusion analysis further identifies Hindi–Urdu and the Sadri–Chhattisgarhi–Surgujia cluster as persistent hard cases due to near-identical spoken forms and high phonological overlap; this supports the interpretation that such clusters are likely to be hard across corpora, including Kathbath (Basu et al., 8 Jun 2026).

7. Access, limitations, and interpretive cautions

Kathbath is publicly connected to two benchmark ecosystems. IndicSUPERB states that code, datasets, and models are available at https://github.com/AI4Bharat/indicSUPERB, while Vistaar states that datasets, code, and models are open-sourced at https://github.com/AI4Bharat/vistaar. The papers also clarify that contributors gave consent and that the data is intended for research and model building.

Several limits on interpretation are explicit. The IndicSUPERB paper does not specify audio sampling rate, file formats, or device specifications in the paper text, and its specific license text is not included there. The Vistaar paper does not provide Kathbath speaker counts, utterance counts, sampling rate/bit depth, train/dev/test splits, transcription conventions, text normalization, or license terms. The cross-corpus LID study does not provide Kathbath-internal details such as recording setups or speaker demographics, does not enumerate the 11 Kathbath languages used in that evaluation, and reports macro accuracy only, without confidence intervals, standard deviations, or per-language Kathbath breakdowns.

A common source of confusion is Kathbath’s changing experimental role across studies. In IndicSUPERB, it is the underlying source corpus from which multiple supervised speech benchmarks are constructed. In Vistaar, it is both an evaluation benchmark and a component of Vistaar-Train. In the later LID study, however, Kathbath is not used for training at all; it is treated strictly as an out-of-domain benchmark relative to Vaani. This distinction is essential when comparing reported numbers across ASR, SLU, and cross-corpus LID settings (Javed et al., 2022, Bhogale et al., 2023, Basu et al., 8 Jun 2026).