@inproceedings{raihan-etal-2026-one,
title = "One Head, Three Transcripts: A Controlled Oracle-to-{ASR} Audit of {B}angla Speech{--}Text Sentiment Analysis",
author = "Raihan, Md and
Rajak, Nilam and
Ghosh, Kuntal and
Dhorali, Sathyanarayanan",
editor = "Ponnusamy, Rahul and
Kumaresan, Prasanna Kumar and
Rajiakodi, Saranya and
B, Premjith and
MC, Shunmuga Priya and
Murugappan, Abirami and
G, Jyothish Lal and
V, Sowmya and
Chakravarthi, Bharathi Raja",
booktitle = "Proceedings of the Fifth Workshop on Multimodal Machine Learning in Low-resource Languages",
month = nov,
year = "2026",
address = "Hengqin, China",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.mmlow-main.4/",
pages = "40--54",
ISBN = "979-8-89176-445-3",
abstract = "Speech sentiment systems are routinely benchmarked with gold transcripts, but a deployed cascade must read whatever its recognizer produces. We audit that gap on BanglaMUSE, a Bangla read-speech sentiment corpus, using sentence-grouped cross-validation and a controlled substitution design in which one trained classifier is evaluated three times while only the test-time transcript changes. Replacing gold transcripts with ASR output costs 0.109 macro-F1 with Bengali XLS-R CTC and 0.341 with our Whisper large-v3 configuration, a high-error stress case. Retraining the classifier on matched recognizer output recovers 57{--}81{\%} of the gap, but re-selecting only the decision threshold on ASR validation data, with weights and probabilities unchanged, already recovers 28{--}49{\%}. Adding the speech embedding improves over ASR text, yet replacing it with acquisition/nuisance features gives a similar gain for gold-trained heads. These features alone predict sentiment at 0.669 macro-F1 (group-permutation null mean 0.495, p = 0.001), and a speech-only classifier that reaches 0.940 falls to 0.621 when both speaker and sentence are unseen. These speech-only and fusion results therefore do not provide clean evidence of prosodic sentiment cues. Low-resource speech{--}text benchmarks should report oracle and deployable conditions, separate threshold effects from retraining, and test acquisition variables before interpreting audio."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="raihan-etal-2026-one">
<titleInfo>
<title>One Head, Three Transcripts: A Controlled Oracle-to-ASR Audit of Bangla Speech–Text Sentiment Analysis</title>
</titleInfo>
<name type="personal">
<namePart type="given">Md</namePart>
<namePart type="family">Raihan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nilam</namePart>
<namePart type="family">Rajak</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Kuntal</namePart>
<namePart type="family">Ghosh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sathyanarayanan</namePart>
<namePart type="family">Dhorali</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-11</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the Fifth Workshop on Multimodal Machine Learning in Low-resource Languages</title>
</titleInfo>
<name type="personal">
<namePart type="given">Rahul</namePart>
<namePart type="family">Ponnusamy</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Prasanna</namePart>
<namePart type="given">Kumar</namePart>
<namePart type="family">Kumaresan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Saranya</namePart>
<namePart type="family">Rajiakodi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Premjith</namePart>
<namePart type="family">B</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Shunmuga</namePart>
<namePart type="given">Priya</namePart>
<namePart type="family">MC</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Abirami</namePart>
<namePart type="family">Murugappan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jyothish</namePart>
<namePart type="given">Lal</namePart>
<namePart type="family">G</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sowmya</namePart>
<namePart type="family">V</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bharathi</namePart>
<namePart type="given">Raja</namePart>
<namePart type="family">Chakravarthi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Hengqin, China</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-445-3</identifier>
</relatedItem>
<abstract>Speech sentiment systems are routinely benchmarked with gold transcripts, but a deployed cascade must read whatever its recognizer produces. We audit that gap on BanglaMUSE, a Bangla read-speech sentiment corpus, using sentence-grouped cross-validation and a controlled substitution design in which one trained classifier is evaluated three times while only the test-time transcript changes. Replacing gold transcripts with ASR output costs 0.109 macro-F1 with Bengali XLS-R CTC and 0.341 with our Whisper large-v3 configuration, a high-error stress case. Retraining the classifier on matched recognizer output recovers 57–81% of the gap, but re-selecting only the decision threshold on ASR validation data, with weights and probabilities unchanged, already recovers 28–49%. Adding the speech embedding improves over ASR text, yet replacing it with acquisition/nuisance features gives a similar gain for gold-trained heads. These features alone predict sentiment at 0.669 macro-F1 (group-permutation null mean 0.495, p = 0.001), and a speech-only classifier that reaches 0.940 falls to 0.621 when both speaker and sentence are unseen. These speech-only and fusion results therefore do not provide clean evidence of prosodic sentiment cues. Low-resource speech–text benchmarks should report oracle and deployable conditions, separate threshold effects from retraining, and test acquisition variables before interpreting audio.</abstract>
<identifier type="citekey">raihan-etal-2026-one</identifier>
<location>
<url>https://aclanthology.org/2026.mmlow-main.4/</url>
</location>
<part>
<date>2026-11</date>
<extent unit="page">
<start>40</start>
<end>54</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T One Head, Three Transcripts: A Controlled Oracle-to-ASR Audit of Bangla Speech–Text Sentiment Analysis
%A Raihan, Md
%A Rajak, Nilam
%A Ghosh, Kuntal
%A Dhorali, Sathyanarayanan
%Y Ponnusamy, Rahul
%Y Kumaresan, Prasanna Kumar
%Y Rajiakodi, Saranya
%Y B, Premjith
%Y MC, Shunmuga Priya
%Y Murugappan, Abirami
%Y G, Jyothish Lal
%Y V, Sowmya
%Y Chakravarthi, Bharathi Raja
%S Proceedings of the Fifth Workshop on Multimodal Machine Learning in Low-resource Languages
%D 2026
%8 November
%I Association for Computational Linguistics
%C Hengqin, China
%@ 979-8-89176-445-3
%F raihan-etal-2026-one
%X Speech sentiment systems are routinely benchmarked with gold transcripts, but a deployed cascade must read whatever its recognizer produces. We audit that gap on BanglaMUSE, a Bangla read-speech sentiment corpus, using sentence-grouped cross-validation and a controlled substitution design in which one trained classifier is evaluated three times while only the test-time transcript changes. Replacing gold transcripts with ASR output costs 0.109 macro-F1 with Bengali XLS-R CTC and 0.341 with our Whisper large-v3 configuration, a high-error stress case. Retraining the classifier on matched recognizer output recovers 57–81% of the gap, but re-selecting only the decision threshold on ASR validation data, with weights and probabilities unchanged, already recovers 28–49%. Adding the speech embedding improves over ASR text, yet replacing it with acquisition/nuisance features gives a similar gain for gold-trained heads. These features alone predict sentiment at 0.669 macro-F1 (group-permutation null mean 0.495, p = 0.001), and a speech-only classifier that reaches 0.940 falls to 0.621 when both speaker and sentence are unseen. These speech-only and fusion results therefore do not provide clean evidence of prosodic sentiment cues. Low-resource speech–text benchmarks should report oracle and deployable conditions, separate threshold effects from retraining, and test acquisition variables before interpreting audio.
%U https://aclanthology.org/2026.mmlow-main.4/
%P 40-54
Markdown (Informal)
[One Head, Three Transcripts: A Controlled Oracle-to-ASR Audit of Bangla Speech–Text Sentiment Analysis](https://aclanthology.org/2026.mmlow-main.4/) (Raihan et al., MMLow 2026)
ACL