Switchboard-1 Release 2

~2,400 two-sided spontaneous English telephone conversations among 543 US speakers (~260 h), separated into two channels.

• Task: asr, speaker
• Languages: en
• Hours: 260
• Domain: conversational telephone
• License: see source terms
• Homepage: https://catalog.ldc.upenn.edu/LDC97S62

Recommendation

A classic benchmark for conversational telephone ASR (Hub5/SWBD) and speaker recognition; choose it for spontaneous narrowband (8 kHz) two-party dialogue or to compare against decades of published WER. It is a paid LDC corpus, and telephone-bandwidth audio differs from modern wideband data.

Getting the data

Obtain from the dataset homepage.

Requires an LDC license / paid membership (catalog ID LDC97S62); not openly downloadable.

Suggested processing

A recommended VoxKitchen pipeline ships in the repository at voxkitchen/templates/pipelines/asr-training-data.yaml — run it with vkit docker run.