IIT Indore's HGAST Fixes Gender Errors in English-to-Hindi, Marathi and Gujarati Speech Translation

When a woman says "I went to the market" in English, a Hindi translation has to make a choice English never asks for: main bazaar gayi (feminine) or main bazaar gaya (masculine). Machine translation systems, trained on data where masculine forms dominate, frequently pick the wrong one. The result is a translation that quietly changes who the speaker is.

Researchers at the Indian Institute of Technology (IIT) Indore say they have built a practical fix. Their tool, Hierarchical Gender Arbitration for Speech Translation (HGAST), detects and corrects gender errors when English speech is translated into Hindi, Marathi and Gujarati — and it does so without retraining the underlying translation model.

The problem: gender that English hides

English first-person sentences rarely mark gender. Hindi, Marathi and Gujarati do, often through verb endings. A speech-translation system therefore has to infer information that is not present in the English words themselves.

Most systems infer it badly. When a woman speaks in English and her words are translated into these languages, AI systems often default to masculine verb forms, effectively erasing the female speaker from the output. For voice assistants, subtitling, customer-service bots and public-service applications serving Indian users, that is both an accuracy problem and a fairness problem.

How HGAST works

HGAST is designed as a correction layer that sits on top of existing translation systems rather than replacing them. According to the IIT Indore team, the tool:

• listens to the speaker's voice to identify the likely gender of the speaker;
• analyses the sentence structure of the translation to find which words — particularly verbs — carry grammatical gender;
• arbitrates and corrects the gendered forms so they match the speaker.

The key engineering choice is that HGAST does not require costly, time-consuming retraining of the translation models it works with. It can be integrated into existing pipelines, which lowers the barrier for companies and public platforms that already run production translation systems.

The researchers tested HGAST on English-to-Hindi, English-to-Marathi and English-to-Gujarati translation, using three different existing translation systems as the base models.

Who built it

The work was led by Dr Chandresh Kumar Maurya, Associate Professor in the Department of Computer Science and Engineering at IIT Indore, together with Dr Mahendra Gupta of Government College of Engineering, Anuppur, and intern Anjil Kumar Raj. Maurya has also co-authored a review of gender bias in spoken-language translation, placing HGAST within a longer research line on the problem.

Why a bolt-on approach matters

Large translation models are expensive to retrain, and many Indian developers build on models they do not control. A modular fix that can be attached to any system is easier to adopt than one that requires rebuilding the core. It also allows the correction step to be audited and improved independently.

There are open questions a correction layer cannot fully resolve. Inferring gender from voice is an imperfect proxy, and a production system will need to handle speakers whose voice does not match their gender identity, as well as sentences where the grammatical gender belongs to someone other than the speaker. Those are design choices for deployers, and they deserve explicit handling rather than silent defaults.

The bigger picture

India is investing heavily in Indian-language AI, from the government's Bhashini language-technology platform to foundation models under the IndiaAI Mission. Speech translation is central to that effort because it lets citizens who do not read English use digital services by voice.

Tools like HGAST address a layer of quality that raw benchmark scores often miss: whether a translation respects who is speaking. As Indian-language voice interfaces move into banking, healthcare and government services, getting a speaker's gender right is not a cosmetic detail. It is part of whether the technology treats the people using it accurately — and that makes this compact piece of research from Indore relevant well beyond the three languages it was tested on.

Sources

• ThePrint — IIT Indore researchers develop tool to ensure fair gender representation in speech translation: https://theprint.in/india/iit-indore-researchers-develop-tool-to-ensure-fair-gender-representation-in-speech-translation/3056863/
• Free Press Journal — IIT Indore Develops AI Tool HGAST To Fix Gender Bias In Speech Translation: https://www.freepressjournal.in/indore/iit-indore-develops-ai-tool-hgast-to-fix-gender-bias-in-speech-translation-from-english-to-hindi-marathi-and-gujarati
• Mediawala — IIT इंदौर के शोधकर्ताओं ने बनाया जेंडर करेक्शन टूल: https://mediawala.in/iit-indore-researchers-develop-gender-correction-tool-ai-translations-will-no-longer-make-gender-based-errors/
• SSRN — Gender Bias in Spoken Language Translation: A Review (Jamal, Maurya): https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6795059