IndicST: Indian Multilingual Translation Corpus For Evaluating Speech Large Language Models

Introduction

IndicST, a new dataset tailored for training and evaluating Speech LLMs for AST tasks (including ASR), featuring meticulously curated, automatically, and manually verified synthetic data. The dataset offers 10.8k hrs of training data and 1.13k hrs of evaluation data.

Use-Cases

ASR (Speech-to-Text)

• Transcribing Indic languages
• Handling accents and noisy environments
• Supporting low-resource language ASR

Automatic Speech Translation (AST)

• Speech-to-speech and speech-to-text translation
• Real-time multilingual communication

Dataset Details

Training data: We utilized ASR data from 14 open-source datasets available publicly, collectively 10.8k hrs spread over nine languages, and Table 1 provides more details. Each dataset consists of input speech audio along with transcription. To synthetically generate the translation for input speech audio and transcription, we used IndicTrans2 tool. we consider two translation directions: one-to-many, where English (source) transcription is translated to text in 8 Indian languages (target), represented as en → X, and many-to-one, where transcription in 8 Indian languages (source) is translated to English (target), represented as X →en.

• One-to-many: English (source) transcription is translated to text in 8 Indian languages (target).
• Many-to-one: Transcription in 8 Indian languages (source) is translated to English (target).

Table1: Summary of ASR Datasets for various Indian Languages used for training Speech LLM. The Duration is mentioned in K Hrs.

Datasets	en	hi	mr	gu	bn	ta	te	ml	kn	Duration (k hrs)
Spring Labs	✔	✔	✘	✘	✘	✘	✘	✘	✘	2.2
Common accent	✔	✘	✘	✘	✘	✘	✘	✘	✘	0.01
MUCS	✘	✔	✔	✘	✔	✘	✘	✘	✘	0.22
CMU	✘	✔	✔	✔	✔	✔	✔	✘	✔	0.06
CommonVoice	✘	✔	✔	✘	✔	✔	✔	✔	✘	1.6
Gramavaani	✘	✔	✘	✘	✘	✘	✘	✘	✘	0.095
Vaani	✘	✔	✔	✔	✔	✔	✔	✘	✔	0.074
Lahaja	✘	✔	✘	✘	✘	✘	✘	✘	✘	0.011
Shrutilipi	✘	✔	✔	✔	✔	✔	✔	✔	✔	5.319
Google Corpus	✘	✘	✔	✔	✘	✔	✔	✔	✔	0.034
Google Fleurs	✘	✘	✔	✔	✔	✔	✔	✔	✔	0.087
Microsoft Speech Corpus	✘	✘	✘	✔	✘	✔	✔	✘	✘	0.12
IISc MILE	✘	✘	✘	✘	✘	✔	✘	✘	✔	0.45
IndicVoices	✘	✔	✔	✔	✔	✔	✔	✔	✔	0.52
Total Duration	1.4	3	1.1	0.5	1.7	1.4	0.5	0.4	0.8	10.8k hrs

The table includes indicators for language availability in datasets, where a check mark (✔) represents availability and a cross (✘) indicates the absence of support for that language.

Test set: For evaluation, we created a test set for Two scenarios

• Input Speech audio available: We used the Kathbath ASR dataset for this scenario to get the X → en translation pair (more details in Table 2) and the Svarah dataset en → X translation pair.
• No input speech audio is available: For this case, we used the AI4Bharat Conv text-to-text translation dataset and speech audio for the source text pair generated using the TTS model. The duration of this test set is available in Table III. More details about this dataset can be found in IndicST paper.

Table 2: Language-wise duration (hrs) of audios in Kathbath.

Language	Duration (hrs)
Hindi (hi)	137.1
Marathi (mr)	166.5
Gujarati (gu)	116.2
Bengali (bn)	104.2
Tamil (ta)	166.3
Telugu (te)	139.2
Malayalam (ml)	132.2
Kannada (kn)	149.2

Table 3: Language-wise duration (min) of audio in AI4Bharath

Language	Duration (mins)
English (en)	28.9
Hindi (hi)	36.1
Marathi (mr)	40.0
Gujarati (gu)	36.0
Bengali (bn)	44.3
Tamil (ta)	39.9
Telugu (te)	45.2
Malayalam (ml)	33.1
Kannada (kn)	35.3

Evaluation Results

We have benchmarked the dataset for ASR and AST tasks using audio-llm (whisper + llama based LLM). We use whisper-large-v2 as a baseline for both the tasks. Results are given in Table IV and V, respectively, for ASR and AST tasks.

Table 4: erformance metric with TP1 (ASR-only) across different models on in-domain Generic-ASR and out-of-domain Svarah and Kathbath test sets. All values are in percentage.

Languages	Baseline	Baseline	Baseline	M1 (TP1)	M1 (TP1)	M1 (TP1)	M2 (TP1)	M2 (TP1)	M2 (TP1)
Languages	Generic-ASR	Svarah	Kathbath	Generic-ASR	Svarah	Kathbath	Generic-ASR	Svarah	Kathbath
en	23.3	25.6		17.7	32		16.5	26.4
hi	63.7		44.5	34.3		14.6	27.3		9.9
mr	99.7		91	29.5		31.9	24.2		29.7
gu	109.4		109.9	56.3		34.2	41.3		25.9
bn	116.6		110.9	69.4		26.8	63.2		26.9
ta	66.6		59.1	37.1		39.3	38		34.6
te	111.3		112.7	75.4		51.1	68.5		37.1
ml	111.7		117.5	47.6		47.2	47.4		46.6
kn	87.7		82.4	56.9		44.2	42.1		30.4

Table 5: Performance metric (BLEU) with TP2 (AST-only) and TP3 (ASR + AST) across different models on in-domain Generic-AST and out-of-domain Svarah, Kathbath, and AI4Bharat test sets. All values are in percentage.

5a. En -> X

Models	Datasets	en→hi	en→mr	en→gu	en→bn	en→ta	en→te	en→ml	en→kn
Baseline	Generic-AST
	Svarah
	Kathbath
	AI4B
M1 (TP2)	Generic-AST	30.2	19.9	25.1	24.4	18.5	19	16.7	18.8
	Svarah	20.9	10.6	14.9	14.5	7.9	10.2	7.4	11.5
	Kathbath
	AI4B	8.8	3.8	7.2	5.3	0.9	1.9	0.6	0.8
M2 (TP2)	Generic-AST	35.6	22.1	29	27.8	21.6	25	20	23.9
	Svarah	28.9	15.1	17.7	19.2	11	14.2	10.6	11
	Kathbath
	AI4B	13.4	6.9	9.5	6.3	1.6	2.1	1.2	1.2
M2 (TP3)	Generic-AST	37	22.6	30.8	28.6	23	25.4	20.6	23.7
	Svarah	23.9	14.7	19.3	18.9	11.8	14.5	10.1	15.2
	Kathbath
	AI4B	14.9	7.3	11.7	8.7	1.6	2.9	1.2	1.3

5b. X --> En

Models	Datasets	hi→en	mr→en	gu→en	bn→en	ta→en	te→en	ml→en	kn→en
Baseline	Generic-AST	16.9	13.1	10.7	7.7	11	7.7	11.9	8.1
	Svarah
	Kathbath	28.1	13.9	16.8	11.8	11.1	12.8	17.6	10.1
	AI4B	28.8	17.1	19.3	19.7	14.5	17.1	15.7	12.7
M1 (TP2)	Generic-AST	29.2	32.4	30	13	24.2	14.6	29	23.8
	Kathbath	36.6	22.3	25.3	20.8	17.7	19	22	15.9
	AI4B	26.2	18.9	19.5	21.4	14.7	16.3	15.9	12.1
M2 (TP2)	Generic-AST	31	32	30.3	14.7	24.6	15	29.6	24.2
	Kathbath	37.2	23.9	25.1	20.6	17.2	19.1	22.4	16.8
	AI4B	26.7	19.2	19.4	22.1	14.7	17.4	16	13
M2 (TP3)	Generic-AST	30.2	33	32.3	15.4	24.4	16.2	30.5	26.2
	Kathbath	38	24.2	25.6	22.3	18.4	20.2	22.5	17.3
	AI4B	26.1	19.6	18.8	21.2	14	17.1	16.5	12.9

Dataset Download

To download the dataset, visit the IndicST Hugging Face Repo:

• IndicST Dataset on Hugging Face

How to Use and Run

To use this dataset in your project, you can load it using a custom data loading script or directly access the files if integrated with a library that supports JSON. Example usage in Python:

import json

def load_dataset(file_path):
    with open(file_path, 'r') as file:
        data = json.load(file)
    return data

# Load the training data
train_data = load_dataset('path/to/ast/train.json')

License

This code repository and the model weights are licensed under the Krutrim Community License.

Citation

@inproceedings{
  sanket2025IndicST,
  title={{IndicST}: Indian Multilingual Translation Corpus For Evaluating Speech Large Language Models},
  author={Sanket Shah, Kavya Ranjan Saxena, Kancharana Manideep Bharadwaj, Sharath Adavanne, Nagaraj Adiga},
  booktitle={Proc. ICASSP},
  year={2025},
}

Contact

Contributions are welcome! If you have any improvements or suggestions, feel free to submit a pull request on GitHub.