DialectalSpeech-ICL

A speech-recognition dataset of African American English (AAE) utterances spanning multiple regional varieties. Each record provides an audio clip, its verbatim reference transcript, and speaker/region metadata intended for evaluating ASR and in-context-learning approaches on dialectal, low-resource speech.

This release is a stratified sample of utterances drawn across all regional collections.

Dataset Structure

Split	Utterances
test	294

Composition by regional collection:

Collection	Region	Utterances
ATL	Atlanta, GA	43
DCA	Washington, D.C.	24
DCB	Washington, D.C.	49
DTA	Detroit, MI	23
LES	Lower East Side, NYC	28
PRV	Princeville, NC	27
ROC	Rochester, NY	48
VLD	Valdosta, GA	52
Total		294

Files

.
├── data.parquet                      # all rows, with embedded audio (HF Audio feature)
├── transcripts.jsonl                 # one row per utterance (metadata + relative audio path)
└── data/
    └── audio/<collection>/*.wav      # the audio clips (16-bit PCM, 44.1 kHz)

Schema

Field	Type	Description
id	string	Stable unique id (<file_base>_l<line>)
audio	Audio	The utterance waveform (decodes to array + sampling rate)
text	string	Verbatim reference transcript
collection	string	Regional collection: ATL, DCA, DCB, DTA, LES, PRV, ROC, VLD
speaker	string	Speaker identifier
file_base	string	Source interview file
duration	float	Clip duration (seconds)
aae_features	string	JSON of detected AAE feature flags (e.g. habitual_be, copula_deletion)
aae_feature_count	int	Number of AAE features detected in the utterance

Usage

from datasets import load_dataset

ds = load_dataset("CentificAIResearch/DialectalSpeech-ICL", split="test")
row = ds[0]
print(row["text"])
audio = row["audio"]          # {"array": ..., "sampling_rate": 44100}

Source & Attribution

Derived from the Corpus of Regional African American Language (CORAAL) — Kendall, Tyler and Charlie Farrington, The Corpus of Regional African American Language, The Online Resources for African American Language (ORAAL) Project. https://oraal.uoregon.edu/coraal

Please cite CORAAL when using this dataset.

License

Released under CC-BY-4.0-NC-SA (CC BY-NC-SA 4.0), matching the CORAAL source license: attribution required, non-commercial use only, derivatives must be shared under the same license. For research and educational purposes only.