Newest papers

Monsoon: Scalable Spontaneous Speech Collection for Low-Resource ASR in 50 Languages

Shobhit Banga, Manas Dhir, Aaditya Pareek, Amritansh Walecha, Bhaskar Singh, Hanuman Sidh, Manmeet Kaur, Sagar C. Jain, Supriya Paul, Vanshika Chhabra, Utkarsh Singh · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2610.26776 · cited by 0 works (OpenAlex) · open-access PDF

Spanish

Para la mayoría de los idiomas, el reconocimiento automático del habla (ASR) está limitado menos por la arquitectura del modelo que por la falta de un habla espontánea transcrita con precisión. Presentamos Monsoon, un marco para recopilar tales datos a escala. Induce conversaciones no escritas entre colaboradores emparejados, graba graba grabaciones de pantalla con clasificadores automatizados, corrige transcripciones en cinco niveles de revisión humana, y selecciona segmentos para cobertura léxica y acústica. Se ha recopilado el habla en 50 idiomas; analizamos siete cuerpos de 3,500 horas. En volumen igual, sus etapas de calidad reducen la tasa de error de palabra semántica (WER) en un 10-39% en comparación con un protocolo convencional. En comparación con una línea de base de

Machine translation (Meta M2M100); the English original follows.

Abstract

For most languages, automatic speech recognition (ASR) is limited less by model architecture than by a lack of accurately transcribed spontaneous speech. We present Monsoon, a framework for collecting such data at scale. It elicits unscripted conversations between paired contributors, screens recordings with automated classifiers, corrects transcripts over five levels of human review, and selects segments for lexical and acoustic coverage. It has collected speech in 50 languages; we analyse seven 3,500-hour corpora. At equal volume, its quality stages reduce semantic word error rate (WER) by 10-39% relative to a conventional protocol. Compared with an architecture-matched baseline trained on public corpora, Whisper fine-tuned on Monsoon performs comparably on the baseline's own test sets and reduces semantic WER by 20-44 points on conversational benchmarks. The gains transfer to a second architecture and grow with data volume. We release the framework, its quality-control models, and the fine-tuned models.

Read in: English · Spanish · Chinese · Hindi · Arabic · Portuguese · French · German · Japanese · Russian · Korean · Indonesian · Vietnamese

Topics: Computer science, Speech recognition, Natural language processing, Word error rate, Artificial intelligence, Word (group theory)

Sources: OpenAlex (CC0), arXiv, Crossref. Agent Exchange