SSMT-Benchmark-English-ILs is a benchmark dataset and evaluation suite designed for Speech-to-Speech Machine Translation (SSMT) research involving English and low-resource Indian languages (ILs).

It provides:

Gold reference translations
ASR transcripts
Evaluation scripts
Standardized test sets
Reproducible benchmarking setup

The goal is to enable fair, consistent, and comparable evaluation of cascaded or end-to-end speech translation systems for Indic languages.

✨ Motivation

Most speech translation benchmarks focus on high-resource or European language pairs. Indian languages remain under-represented despite:

Large speaker populations
High linguistic diversity
Limited speech resources
Frequent disfluencies in spoken data
Domain variability (education, governance, public services)

This benchmark helps address these gaps by providing carefully curated evaluation data for English → Indic speech translation.

It is particularly useful for:

ASR → MT → TTS pipelines
End-to-end Speech Translation models
Low-resource MT research
Indic language modeling
Disfluency-aware translation systems

📂 Repository Structure


SSMT-Benchmark-English-ILs/
│
├── asr/                # ASR transcripts or system outputs
├── gold/               # Gold reference translations
├── eval_scripts/       # BLEU / COMET / ChrF evaluation scripts
├── data_stats/         # Dataset statistics and metadata
├── LICENSE
└── README.md

Folder Description

Folder	Description
`gold/`	Human translated references for evaluation
`asr/`	Example or placeholder ASR outputs / system hypotheses
`eval_scripts/`	Metric computation scripts
`data_stats/`	Corpus statistics and metadata

🌍 Supported Languages

The benchmark covers English → multiple Indian languages, including:

Hindi (hin)
Bengali (ben)
Marathi (mar)
Tamil (tam)
Telugu (tel)
Gujarati (guj)
Kannada (kan)
Malayalam (mal)
(additional languages may be included)

Each language directory contains aligned files:


source.en
reference.xx

🚀 Quick Start

1. Clone the repository

git clone https://github.com/vmujadia/SSMT-Benchmark-English-ILs.git
cd SSMT-Benchmark-English-ILs

2. Generate predictions

Run your Speech Translation pipeline:

English Speech
   ↓
ASR
   ↓
MT
   ↓
(Optional TTS)
   ↓
Translated text

Save predictions in plain text:

my_outputs.txt

(one sentence per line, aligned with references)

3. Evaluate

BLEU (SacréBLEU)

pip install sacrebleu

sacrebleu gold/reference.hin -i my_outputs.txt -m bleu chrf

COMET

pip install unbabel-comet

comet-score -s gold/source.en -t my_outputs.txt -r gold/reference.hin

Using provided script (example)

python eval_scripts/calc_bleu.py \
  --hyp my_outputs.txt \
  --ref gold/reference.hin

📊 Recommended Metrics

We suggest reporting:

Metric	Why
BLEU	Standard MT metric
ChrF	Better for morphologically rich Indic languages
COMET	Neural, semantic quality metric

For speech translation, combining BLEU + COMET gives more reliable evaluation.

🔬 Example Research Use Cases

This benchmark can be used for:

Cascaded SSMT (ASR→MT→TTS)
End-to-end Speech Translation
Disfluency removal before MT
Low-resource fine-tuning
Domain adaptation
Indic multilingual models
Benchmarking Whisper + MT pipelines
Indic LLM-based translation

📚 Citation

If you use this dataset or benchmark, please cite:

Mujadia et al., 2025 — Language Resources and Evaluation

Disfluency processing for cascaded speech translation involving English and Indian languages

BibTeX

@article{Mujadia2025Disfluency,
  title={Disfluency processing for cascaded speech translation involving English and Indian languages},
  author={Mujadia, Vandan and Mishra, Pruthwik and Sharma, Dipti Misra},
  journal={Language Resources and Evaluation},
  volume={59},
  pages={2653--2686},
  year={2025},
  publisher={Springer}
}

Paper link: https://link.springer.com/article/10.1007/s10579-025-09818-3