Research

Paper

AI LLM March 06, 2026

Doctor or Patient? Synergizing Diarization and ASR for Code-Switched Hinglish Medical Conditions Extraction

Authors

Séverin Baroudi, Yanis Labrak, Shashi Kumar, Joonas Kalda, Sergio Burdisso, Pawel Cyrta, Juan Ignacio Alvarez-Trejos, Petr Motlicek, Hervé Bredin, Ricard Marxer

Abstract

Extracting patient medical conditions from code-switched clinical spoken dialogues is challenging due to rapid turn-taking and highly overlapped speech. We present a robust system evaluated on the DISPLACE-M dataset of real-world Hinglish medical conversations. We propose an End-to-End Neural Diarization with Vector Clustering approach (EEND-VC) to accurately resolve dense and speaker overlaps in Doctor-Patient Conversations (DoPaCo). For transcription, we adapt a Qwen3 ASR model via domain-specific fine-tuning, Devanagari script normalization, and dialogue-level LLM error correction, achieving an 18.59% tcpWER. We benchmark open and proprietary LLMs on medical condition extraction, comparing our text-based cascade system against a multimodal End-to-End (E2E) audio framework. While proprietary E2E models set the performance ceiling, our open cascaded architecture is highly competitive, as it achieved first place out of 25 participants in the DISPLACE-M challenge. All implementations are publicly released.

Metadata

arXiv ID: 2603.06373
Provider: ARXIV
Primary Category: eess.AS
Published: 2026-03-06
Fetched: 2026-03-09 06:05

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.06373v1</id>\n    <title>Doctor or Patient? Synergizing Diarization and ASR for Code-Switched Hinglish Medical Conditions Extraction</title>\n    <updated>2026-03-06T15:24:52Z</updated>\n    <link href='https://arxiv.org/abs/2603.06373v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.06373v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Extracting patient medical conditions from code-switched clinical spoken dialogues is challenging due to rapid turn-taking and highly overlapped speech. We present a robust system evaluated on the DISPLACE-M dataset of real-world Hinglish medical conversations. We propose an End-to-End Neural Diarization with Vector Clustering approach (EEND-VC) to accurately resolve dense and speaker overlaps in Doctor-Patient Conversations (DoPaCo). For transcription, we adapt a Qwen3 ASR model via domain-specific fine-tuning, Devanagari script normalization, and dialogue-level LLM error correction, achieving an 18.59% tcpWER. We benchmark open and proprietary LLMs on medical condition extraction, comparing our text-based cascade system against a multimodal End-to-End (E2E) audio framework. While proprietary E2E models set the performance ceiling, our open cascaded architecture is highly competitive, as it achieved first place out of 25 participants in the DISPLACE-M challenge. All implementations are publicly released.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='eess.AS'/>\n    <published>2026-03-06T15:24:52Z</published>\n    <arxiv:comment>Submitted for review at Interspeech 2026</arxiv:comment>\n    <arxiv:primary_category term='eess.AS'/>\n    <author>\n      <name>Séverin Baroudi</name>\n    </author>\n    <author>\n      <name>Yanis Labrak</name>\n    </author>\n    <author>\n      <name>Shashi Kumar</name>\n    </author>\n    <author>\n      <name>Joonas Kalda</name>\n    </author>\n    <author>\n      <name>Sergio Burdisso</name>\n    </author>\n    <author>\n      <name>Pawel Cyrta</name>\n    </author>\n    <author>\n      <name>Juan Ignacio Alvarez-Trejos</name>\n    </author>\n    <author>\n      <name>Petr Motlicek</name>\n    </author>\n    <author>\n      <name>Hervé Bredin</name>\n    </author>\n    <author>\n      <name>Ricard Marxer</name>\n    </author>\n  </entry>"
}