@inproceedings{rai-etal-2026-railvaani,
title = "{R}ail{V}aani: An Optimized Bilingual {ASR} Pipeline for Entity-Rich Railway Announcements",
author = "Rai, Anand Kumar and
Desai, Swar and
Mittal, Shshank and
Mukherjee, Animesh",
editor = "Agrawal, Sweta and
Abad, Alberto",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-demos.51/",
pages = "528--536",
ISBN = "979-8-89176-461-3",
abstract = "Automatic Speech Recognition (ASR) for railway announcements is challenging due to noisy environments and dense, entity-rich content. We introduce RailVaani, a low-latency bilingual (Hindi-English) ASR pipeline for Indian railway public address (PA) announcements, benchmarked on RailKosh {--} the first bilingual annotated corpus of real railway PA audio (1,513 clips: 1,000 English, 513 Hindi; 53 stations; gold transcripts and entity labels). We evaluate ASR baselines and their LoRA fine-tuned variants with database-aware (DB-aware) post-processing (PP), using WER and task score (mean of train number accuracy, platform accuracy, and announcement-type macro-F1), as task score is the more passenger-relevant metric for this domain. Our key findings: (i) on English PA, our deployed system (Whisper Small + LoRA + PP) achieves 69.8{\%} task score at a real-time factor (RTF) of 0.75; a larger variant (Qwen3-ASR-0.6B + LoRA + PP) matches Google STT (79.7{\%} vs. 79.3{\%}) at RTF=1.37; (ii) on Hindi PA, our deployed system (collabora/whisper-small-hindi + LoRA + PP) achieves 60.7{\%} task score at RTF=0.83, surpassing Whisper Large v3 (43.5{\%}) with a 6x smaller model; and (iii) a user study (N=63; SUS-style median=90.0) indicates strong perceived usability. Both deployed systems run faster than real-time on CPU, enabling deployment without GPU infrastructure. RailVaani, RailKosh, and the RailKosh annotation toolkit are fully open-source."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="rai-etal-2026-railvaani">
<titleInfo>
<title>RailVaani: An Optimized Bilingual ASR Pipeline for Entity-Rich Railway Announcements</title>
</titleInfo>
<name type="personal">
<namePart type="given">Anand</namePart>
<namePart type="given">Kumar</namePart>
<namePart type="family">Rai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Swar</namePart>
<namePart type="family">Desai</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shshank</namePart>
<namePart type="family">Mittal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Animesh</namePart>
<namePart type="family">Mukherjee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations</title>
</titleInfo>
<name type="personal">
<namePart type="given">Sweta</namePart>
<namePart type="family">Agrawal</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alberto</namePart>
<namePart type="family">Abad</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-461-3</identifier>
</relatedItem>
<abstract>Automatic Speech Recognition (ASR) for railway announcements is challenging due to noisy environments and dense, entity-rich content. We introduce RailVaani, a low-latency bilingual (Hindi-English) ASR pipeline for Indian railway public address (PA) announcements, benchmarked on RailKosh – the first bilingual annotated corpus of real railway PA audio (1,513 clips: 1,000 English, 513 Hindi; 53 stations; gold transcripts and entity labels). We evaluate ASR baselines and their LoRA fine-tuned variants with database-aware (DB-aware) post-processing (PP), using WER and task score (mean of train number accuracy, platform accuracy, and announcement-type macro-F1), as task score is the more passenger-relevant metric for this domain. Our key findings: (i) on English PA, our deployed system (Whisper Small + LoRA + PP) achieves 69.8% task score at a real-time factor (RTF) of 0.75; a larger variant (Qwen3-ASR-0.6B + LoRA + PP) matches Google STT (79.7% vs. 79.3%) at RTF=1.37; (ii) on Hindi PA, our deployed system (collabora/whisper-small-hindi + LoRA + PP) achieves 60.7% task score at RTF=0.83, surpassing Whisper Large v3 (43.5%) with a 6x smaller model; and (iii) a user study (N=63; SUS-style median=90.0) indicates strong perceived usability. Both deployed systems run faster than real-time on CPU, enabling deployment without GPU infrastructure. RailVaani, RailKosh, and the RailKosh annotation toolkit are fully open-source.</abstract>
<identifier type="citekey">rai-etal-2026-railvaani</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-demos.51/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>528</start>
<end>536</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T RailVaani: An Optimized Bilingual ASR Pipeline for Entity-Rich Railway Announcements
%A Rai, Anand Kumar
%A Desai, Swar
%A Mittal, Shshank
%A Mukherjee, Animesh
%Y Agrawal, Sweta
%Y Abad, Alberto
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-461-3
%F rai-etal-2026-railvaani
%X Automatic Speech Recognition (ASR) for railway announcements is challenging due to noisy environments and dense, entity-rich content. We introduce RailVaani, a low-latency bilingual (Hindi-English) ASR pipeline for Indian railway public address (PA) announcements, benchmarked on RailKosh – the first bilingual annotated corpus of real railway PA audio (1,513 clips: 1,000 English, 513 Hindi; 53 stations; gold transcripts and entity labels). We evaluate ASR baselines and their LoRA fine-tuned variants with database-aware (DB-aware) post-processing (PP), using WER and task score (mean of train number accuracy, platform accuracy, and announcement-type macro-F1), as task score is the more passenger-relevant metric for this domain. Our key findings: (i) on English PA, our deployed system (Whisper Small + LoRA + PP) achieves 69.8% task score at a real-time factor (RTF) of 0.75; a larger variant (Qwen3-ASR-0.6B + LoRA + PP) matches Google STT (79.7% vs. 79.3%) at RTF=1.37; (ii) on Hindi PA, our deployed system (collabora/whisper-small-hindi + LoRA + PP) achieves 60.7% task score at RTF=0.83, surpassing Whisper Large v3 (43.5%) with a 6x smaller model; and (iii) a user study (N=63; SUS-style median=90.0) indicates strong perceived usability. Both deployed systems run faster than real-time on CPU, enabling deployment without GPU infrastructure. RailVaani, RailKosh, and the RailKosh annotation toolkit are fully open-source.
%U https://aclanthology.org/2026.emnlp-demos.51/
%P 528-536
Markdown (Informal)
[RailVaani: An Optimized Bilingual ASR Pipeline for Entity-Rich Railway Announcements](https://aclanthology.org/2026.emnlp-demos.51/) (Rai et al., EMNLP 2026)
ACL