EpaDB: English Pronunciation by Argentinians

Dataset Summary

EpaDB is a speech database intended for research in pronunciation scoring. The corpus includes audios from 50 Spanish speakers (25 males and 25 females) from Argentina reading phrases in English. Each speaker recorded 64 short phrases containing sounds hard to pronounce for this population adding up to ~3.5 hours of speech.

Supported Tasks

• Pronunciation Assessment – predict utterance-level global scores or phoneme-level correct/incorrect
• Phone Recognition - predict phoneme sequences
• Phone-level Error Detection – classify each phone as insertion, deletion, distortion, substitution, or correct.
• Alignment Analysis – leverage MFA timings to study forced alignment quality or to refine pronunciation models.

Languages

• L2 utterances: English
• Speaker L1: Spanish

Dataset Structure

Data Instances

Each JSON entry describes one utterance:

• Phone sequences for reference transcription (reference) and annotators (annot_1, optional annot_2).
• Phone-level labels (label_1, label_2) and derived error_type categories.
• MFA start/end timestamps per phone (start_mfa, end_mfa).
• Per-utterance global scores (global_1, global_2) and propagated speaker levels (level_1, level_2).
• Speaker metadata (speaker_id, gender).
• Audio metadata (duration, sample_rate, wav_path) plus the waveform itself.
• Reference sentence orthographic transcription (transcription).

Data Fields

Field	Type	Description
utt_id	string	Unique utterance identifier (e.g., spkr28_1).
speaker_id	string	Speaker identifier.
sentence_id	string	Reference sentence ID (matches reference_transcriptions.txt).
phone_ids	sequence[string]	Unique phone identifiers per utterance.
reference	sequence[string]	reference phones assigned to match the closer aimed pronunciation by the speaker.
annot_1	sequence[string]	Annotator 1 phones (- marks deletions).
annot_2	sequence[string]	Annotator 3 phones when available, empty otherwise.
label_1	sequence[string]	Annotator 1 phone labels ("1" correct, "0" incorrect).
label_2	sequence[string]	Annotator 3 phone labels when present.
error_type	sequence[string]	Derived categories: correct, insertion, deletion, distortion, substitution.
start_mfa	sequence[float]	Phone start times (seconds).
end_mfa	sequence[float]	Phone end times (seconds).
global_1	float or null	Annotator 1 utterance-level score (1–4).
global_2	float or null	Annotator 3 score when available.
level_1	string or null	Speaker-level proficiency tier from annotator 1 ("A"/"B").
level_2	string or null	Speaker tier from annotator 3.
gender	string or null	Speaker gender ("M"/"F").
duration	float	Utterance duration in seconds (after resampling to 16 kHz).
sample_rate	int	Sample rate in Hz (16,000).
wav_path	string	Waveform filename (<utt_id>.wav).
audio	Audio	Automatically loaded waveform (16 kHz).
transcription	string or null	Reference sentence text.

Data Splits

Split	# Examples
train	1,903
test	1,263

Notes

• When annotator 3 did not label an utterance, related fields (annot_2, label_2, global_2, level_2) are absent or set to null.
• Error types come from simple heuristics contrasting MFA reference phones with annotator 1 labels.
• Waveforms were resampled to 16 kHz using ffmpeg during manifest generation.
• Forced alignments and annotations were merged to produce enriched CSV files per speaker/partition.
• Global scores are averaged per speaker to derive level_* tiers (A if mean ≥ 3, B otherwise).

Licensing

• Audio and annotations: CC BY-NC 4.0 (non-commercial use allowed with attribution).

Citation

@article{vidal2019epadb,
  title   = {EpaDB: a database for development of pronunciation assessment systems},
  author  = {Vidal, Jazmin and Ferrer, Luciana and Brambilla, Leonardo},
  journal = {Proc. Interspeech},
  pages   = {589--593},
  year    = {2019}
}

Usage

Install dependencies and load the dataset:

from datasets import load_dataset

# Local usage before uploading:
ds = load_dataset(
    "epadb_dataset/epadb.py",
    data_dir="/path/to/epadb",  # folder with train.json, test.json, WAV/
    split="train",
)
print(ds)
print(ds[0]["utt_id"], ds[0]["audio"]["sampling_rate"])  # 16000

# After pushing to the Hugging Face Hub:
# ds = load_dataset("JazminVidal/epadb", split="train")

Acknowledgements

The database is an effort of the Speech Lab at the Laboratorio de Inteligencia Artificial Aplicada from the Universidad de Buenos Aires and was partially funded by Google by a Google Latin America Reseach Award in 2018