burkimbia/speech-rejected-aligned

Speech rejected, aligned

Mooré speech segments recovered from clips that our curation pipeline rejected for being too long. The source is burkimbia/speech-rejected. Its card names aligning the long audio as the next step; this dataset is that step.

Key statistics

• Segments: 1,129
• Total duration: about 5.0 hours
• Segment length: 0.5–20.6 s, median 19.1 s
• Origin: 993 segments from JW, 136 from BIBLE.com
• Speakers: 21 speaker labels, all male
• Audio: FLAC, 16 kHz, mono
• Language: Mooré (mos)

How it was made

1. Select: all rows of burkimbia/speech-rejected longer than 30 s, which is 889 clips and 8.7 h of audio.

2. Align: each transcript was forced-aligned to its audio at word level with aadel4/omniASR-CTC-1B-v2 through easyaligner. One clip, whose transcript was too long for its audio, could not be aligned and was skipped.

3. Segment: consecutive words were grouped into segments of about 20 s at most, and a new segment starts at any pause longer than 2 s. The transcripts have no punctuation, so other cuts are made on length, at word boundaries.

4. Filter: a segment is kept only if it passes three checks:

   • align_score >= 0.7;
   • at least 5 characters per second, the same threshold as the curation filter;
   • no word longer than 1.5 s.

   The last two catch a case the score misses: when the end of a transcript is matched to audio that isn't those words, such as silence or music, the words get stretched over it but each still scores high. 195 of 1,324 unique segments fail a check.

5. Deduplicate: speech-rejected contains the same recording several times, ingested from different source shards. One copy of each segment is kept: 2,010 segments in total, 686 of them duplicates.

Code: scripts/modal_align.py in BurkimbIA/speech-data-processing.

Data fields

Field	Description
id	<source clip>_<segment index>
audio	Segment audio, 16 kHz
text	Mooré transcript of the segment, without punctuation
start_time, end_time	Position of the segment in the source clip, in seconds
duration	Segment length in seconds
align_score	Mean word-level alignment confidence, 0.7 to 1. Median 0.90
origin	Where the recording comes from: JW or BIBLE.com
speakers	Speaker label of the kept copy
speakers_all	All speaker labels seen across the copies of this segment, comma-separated
gender	Speaker gender
source_clip	File name of the source row in burkimbia/speech-rejected (its audio.path)
source_shard, source_key	Raw WebDataset shard and sample key of the source row

Known limitations

• Speaker labels are not reliable across copies. Diarization ran separately on each copy of a duplicated recording, so the same voice can have different labels. 292 segments have more than one label in speakers_all. Keep this in mind when building speaker-disjoint train and test splits.
• Speaker and gender diversity is low. All segments are male speech read from religious texts.
• Alignment is only as good as the transcript. A score of 0.7 removes most mismatched segments, but some text at segment edges may not be spoken in the audio. Median speaking rate is 3.0 words/s; a segment far above that likely has extra text.
• Use source_clip to trace a segment back to its source. The id column of burkimbia/speech-rejected is not unique, so it can't be used for that join.

Usage

from datasets import load_dataset

ds = load_dataset("burkimbia/speech-rejected-aligned", split="train")

# Stricter alignment threshold
ds = ds.filter(lambda score: score >= 0.85, input_columns="align_score")

Full alignment output

The Burkimbia team keeps the unfiltered output in the private HF bucket burkimbia/speech-alignments, under speech-rejected-aligned/:

• manifest.jsonl: all 2,010 segments, including rejected ones (with reasons) and duplicates.
• alignments.jsonl: word-level timestamps and scores for every source clip.

Use these to pick another threshold or re-cut segments without re-running alignment.