CoolFace
Datasetpublic

brucemacd/transatlantic-voice-archive

Transatlantic Voice Archive Speech clips with aligned transcripts in the transatlantic (Mid-Atlantic) accent — the clipped, semi-British delivery of 1930s–1960s American newsreel announcers. Built from public-domain Universal Newsreels (1929–1967) on the Internet Archive, intended for finetuning TTS models on the accent. Dataset statistics Clips 1427 Total audio 2.04 h (122.2 min) Average clip 5.14 s Sample rate 22050 Hz, mono WAV Source reels… See the full description on the dataset page: https://huggingface.co/datasets/brucemacd/transatlantic-voice-archive.

sourceHugging Faceotherupdated 3mo agoView on Hugging Face
0likes20downloads
Dataset Card

Transatlantic Voice Archive

Speech clips with aligned transcripts in the transatlantic (Mid-Atlantic) accent — the clipped, semi-British delivery of 1930s–1960s American newsreel announcers. Built from public-domain Universal Newsreels (1929–1967) on the Internet Archive, intended for finetuning TTS models on the accent.

Dataset statistics

Clips1427
Total audio2.04 h (122.2 min)
Average clip5.14 s
Sample rate22050 Hz, mono WAV
Source reels62

Fields

  • —audio — the clip (decoded from wavs/*.wav)
  • —text — transcript as transcribed/parsed (original casing & punctuation)
  • —norm_text — lowercased, normalized transcript for TTS training
  • —source_id — Internet Archive identifier of the source newsreel (https://archive.org/details/<source_id>)

How it was built

  1. 1.Newsreels fetched from the Internet Archive (smallest usable A/V file per item).
  2. 2.Transcripts taken from the item's existing .srt/.vtt subtitles when present, otherwise transcribed with Whisper (faster-whisper, small.en).
  3. 3.Audio cut on subtitle/Whisper segment boundaries into 1–14 s mono clips; non-speech segments (music, junk) filtered heuristically.

Limitations

  • —Transcripts are machine-generated and contain errors; they have not been fully hand-corrected.
  • —Some clips contain background music, crowd noise, or non-narrator speech.
  • —Multiple announcers appear across reels — this is not a single-speaker dataset.
  • —Period recordings: bandwidth-limited, optical-soundtrack audio quality.

Licensing

Universal Newsreels were released into the public domain (Universal did not renew copyrights and donated the films to the U.S. National Archives). Each clip's source_id links back to the source item on the Internet Archive so per-item rights can be verified.