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

Key points

• For most languages, automatic speech recognition (ASR) is limited less by model architecture than by a lack of accurately transcribed spontaneous speech.1
• It has collected speech in 50 languages; we analyse seven 3,500-hour corpora.1
• 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.1
• We release the framework, its quality-control models, and the fine-tuned models.1

Sources (1)

• [1]Monsoon: Scalable Spontaneous Speech Collection for Low-Resource ASR in 50 Languages
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 05:30 PM

  We present Monsoon, a framework for collecting such data at scale.

  For most languages, automatic speech recognition (ASR) is limited less by model architecture than by a lack of accurately transcribed spontaneous speech.

Extractive summary: sentences quoted from the sources.