Hume offers a speech entity dataset for names, addresses and codes

Roughly 800 hours across eight languages come with three types of labels. Access is offered through licensing rather than an open download.

Train for the words a task depends on

Hume introduced the ASR Entity Dataset on November 20 for speech-recognition post-training. The company reports roughly 800 hours of read speech in eight languages and more than 350,000 tagged entities across 28 categories, including names, addresses, phone numbers, emails, dates and codes. The focus is information that scheduling or delivery workflows must preserve accurately, beyond whether a sentence sounds fluent.

Separate recognition from formatting

Each recording has written, spoken and entity-tagged labels, helping developers distinguish recognition errors from formatting mistakes. About 6,000 participants from 49 countries recorded in real homes and offices, with foreign-origin entities included across languages. Access is offered by contacting the team for licensing; the announcement provides neither an open download nor a standard price or guaranteed accuracy gain. Teams should check language coverage and licensing, then split training and testing by speaker to avoid overstating performance through speaker overlap.

via: Hume AI