CALLHOME Spanish: 38 Hours of Unscripted Telephone Conversations

38 hours of speech from 120 unscripted telephone conversations between native Spanish speakers, developed by the Linguistic Data Consortium. This second edition combines original audio and transcripts with updated formats and revised transcriptions conforming to modern guidelines. Calls originated in North America to overseas locations, with participants speaking to family or friends on topics of their choice.

Use Cases

• Train speaker identification models based on unscripted, two-channel telephone audio.
• Develop Spanish speech recognition systems using manually transcribed conversational data.
• Research language identification and dialect analysis based on speaker metadata including accent and region.
• Study conversational dynamics and background noise in telephony applications.

Strengths

• Approximately 38 hours of speech from 120 distinct conversations.
• Includes two versions of transcripts: original and revised with updated annotation standards.
• Audio provided in a modern, lossless FLAC format converted from original recordings.
• Contains detailed metadata on call quality, speaker demographics, and background noise.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The original training/development/test partitioning was removed, which may affect reproducibility of prior studies.
• Data may reflect geographic and social bias inherent to calls placed from North America to overseas contacts.

Provenance

Source

Linguistic Data Consortium (LDC) and Rutgers, The State University of New Jersey.

Collection Method

Manually audited telephone conversations recorded as 8 kHz u-law SPHERE files, later converted to FLAC.

Freshness

Last updated 2026-04-25 08:10:15; freshness should be verified.

Geography

Calls originated in North America and were placed to overseas locations.