Buckeye Corpus - Forced Alignment Benchmark

Segmented utterances from the Buckeye Corpus of Conversational Speech (v2.0) with human-annotated word-level timestamps, prepared for forced alignment benchmarking.

Dataset Stats

Metric	Value
Speakers	20 (s01-s20)
Segments	2,478
Total words	145,762
Total audio	16.1 hours
Audio format	16kHz mono PCM WAV
Avg segment	23.3s
Words per segment	3-123 (avg 59)

Structure

buckeye-forced-alignment-benchmark/
  manifest.json       # metadata + ground truth word timestamps
  audio/
    s0101a_000.wav    # segmented audio clips
    s0101a_001.wav
    ...

manifest.json Format

{
  "dataset": "Buckeye Corpus v2.0",
  "speakers": 20,
  "total_segments": 2478,
  "total_words": 145762,
  "samples": [
    {
      "id": "s0101a_000",
      "speaker": "s01",
      "audio": "audio/s0101a_000.wav",
      "transcript": "okay um i'm lived in columbus...",
      "duration_s": 24.96,
      "num_words": 25,
      "words": [
        {"word": "okay", "start_ms": 100.0, "end_ms": 505.5},
        {"word": "um", "start_ms": 12501.4, "end_ms": 12830.3},
        ...
      ]
    }
  ]
}

Ground Truth

Word timestamps are from Buckeye's human phonetic labeling — hand-corrected by trained annotators. The start_ms and end_ms for each word are derived from the .words annotation files.

Segmentation

Long conversation recordings (5-12 min each) were segmented into utterances at silence/noise boundaries (<SIL>, <IVER>, <VOCNOISE> tokens with gaps > 0.3s). Segments are 2-25 seconds with minimum 3 real words.

Intended Use

Benchmarking forced alignment systems by comparing predicted word timestamps against human ground truth. Primary metric: AAS (Accumulated Average Shift) — mean absolute boundary error in milliseconds.

Citation

Pitt, M.A., Johnson, K., Hume, E., Kiesling, S., & Raymond, W. (2005).
The Buckeye corpus of conversational speech: labeling conventions and a test of transcriber reliability.
Speech Communication, 45, 89-95.

License

The Buckeye Corpus is free for noncommercial use. See buckeyecorpus.osu.edu for terms.