Monsoon framework collects high-quality speech data in 50 languages
Monsoon: Scalable Spontaneous Speech Collection for Low-Resource ASR in 50 Languages
Computation and Language
Summary
Getting computers to understand spontaneous speech well is hard because there aren't enough good examples of real conversations in many languages. The authors present Monsoon, a system that collects natural talk between people, checks the recordings automatically, and improves transcripts with several rounds of human review. They tested Monsoon on 50 languages and found it greatly reduces errors in understanding compared to usual methods. This better data helps improve speech recognition systems across different models and languages.
What this means in practice
- •For speech technology developers: Train speech recognition systems for under-resourced languages using high-quality natural conversation data collected by Monsoon.
- •For multilingual customer service teams: Deploy fine-tuned ASR models from Monsoon data to improve voice-based support handling in multiple low-resource languages.$Commercial implications: Enables better speech recognition products serving diverse customer languages through improved training data and models.
Authors
Shobhit Banga, Manas Dhir, Aaditya Pareek, Amritansh Walecha, Bhaskar Singh, Hanuman Sidh, Manmeet Kaur, Sagar Jain, Supriya Paul, Vanshika Chhabra, Utkarsh Singh
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.