From the source
Hume released the Hume ASR Entity Dataset, a custom post-training dataset for speech recognition models, after testing 45 leading ASR systems that failed on entities such as names, addresses and codes.
- The dataset holds about 800 hours of scripted read speech in eight languages with more than 350,000 tagged entities across 28 categories, covering person names, addresses, phone numbers, account numbers, amounts, dates and spelled-out strings.
- Hume says entity errors derail voice-agent tasks: a wrong email or date of birth can block bank-account verification, a mistaken address can misdirect a delivery, and a mistranscribed passcode can stop someone turning off an alarm.
- About 40% of entities in each language come from another locale, and roughly 6,000 speakers from 49 countries recorded on their own laptops and phones with no noise suppression or level normalization.
- Depending on the language, 5% to 17% of recordings have a signal-to-noise ratio below 15 dB, and another 21% to 30% fall between 15 and 25 dB.
- Each recording carries written, spoken and tagged labels so developers can separate pronunciation from text-normalization errors and build speaker-disjoint splits; Hume licenses the dataset to teams and can collect more hours for any language.






