bbdontcry/en-everyday-conversation-asr

English Everyday-Conversation ASR dataset

34777 clips · 39.996 hours · 16 kHz mono · CC-BY-SA 4.0

Assembled from two commercial-safe (CC-BY-SA 4.0) open corpora: EdAcc (spontaneous dyadic conversation, 11485 clips) and DailyTalk (scripted everyday-life dialogue, 23292 clips). Splits ({'train': 31297, 'dev': 3480}) are conversation-disjoint (no speaker leakage). Per-row metadata (gender, accent, style, l1, ...) is preserved so the set can be re-balanced.

from datasets import load_dataset
ds = load_dataset("<your-username>/<repo>", split="train")
print(ds[0]["text"], ds[0]["audio"]["sampling_rate"])

---

DATASHEET — English Everyday-Conversation ASR set

Motivation

An English everyday-conversation ASR corpus assembled only from commercial-safe (CC-BY-SA 4.0) open sources, for fine-tuning / evaluating speech-to-text on casual, daily-life speech. No telephone-casual corpus (Switchboard/Fisher/CALLHOME) is included because none are commercially licensable.

Composition

• 34777 clips / 39.996 hours total.
• By source: {'edacc': 11485, 'dailytalk': 23292}
• By split (speaker/conversation-disjoint): {'train': 31297, 'dev': 3480}
• Audio: 16 kHz, mono, PCM16 WAV (resampled with soxr).
• Per-clip fields: id, audio, text (normalized), text_raw (verbatim), duration, speaker, conversation, source, style (spontaneous|scripted), license, split_src, split, accent, raw_accent, gender, l1. Metadata is preserved so the set can be re-balanced later (e.g. by gender / accent / style).
• Accent distribution (top): {'American English': 23292, 'Mainstream US English': 1065, 'Southern British English': 901, 'Irish English': 801, 'Nigerian English': 800, 'Indian English': 760, 'Kenyan English': 699, 'Spanish': 694, 'Vietnamese': 595, 'Eastern European': 582, 'Jamaican English': 503, "Don't know": 477, 'Italian': 410, 'Lithuanian': 397, 'Chinese': 381, 'Scottish English': 299, 'Egyptian': 264, 'Catalan': 222, 'Israeli': 207, 'Indonesian English': 184, 'South African English': 156, 'European': 151, 'Indonesian': 139, 'Bulgarian': 124, 'Latin American': 119, 'Romanian': 106, 'Ghanain English': 100, 'Japanese': 86, 'French': 71, 'Dutch': 60}

Collection & preprocessing

Clips come pre-segmented from each source. Processing: downmix→mono, resample→16 kHz, text lowercased keeping apostrophes, event/markup tokens stripped.

• Filtering: keep 1.0s ≤ duration ≤ 20.0s, non-empty text.
• Dropped (transparent, not silent): {'empty_or_marker_text': 401, 'too_short': 6544, 'too_long': 1188}
• Non-speech / markup tokens encountered & removed: {'IGNORE_TIME_SEGMENT_IN_SCORING': 123, '': 639, '': 454, '': 1668, '': 68, '': 160, '': 8, '': 27, '': 5, '': 22, '': 15}

Recommended uses & limitations

• EdAcc is officially an evaluation benchmark (validation/test only). If you evaluate on EdAcc, hold its speakers out using split_src.
• DailyTalk is scripted/acted studio speech (2 voices): good for everyday-life register, not for spontaneity/overlap.
• No literal 'shopping/market field recordings' exist under a commercial-safe licence; 'everyday conversation' here means the casual daily-life register.

Distribution

CC-BY-SA 4.0 — see ATTRIBUTION.md. Keep attributions; redistribute under the same licence.

---

ATTRIBUTION & LICENSE

This dataset is a derivative assembled from the open corpora listed below. Every source is distributed under CC-BY-SA 4.0, so this combined dataset is also released under CC-BY-SA 4.0 (BY = keep the attributions below; SA = redistribute derivatives under the same licence).

https://creativecommons.org/licenses/by-sa/4.0/

Edinburgh International Accents of English Corpus (EdAcc) (source=edacc, 11485 clips)

• HuggingFace: edinburghcstr/edacc
• License: CC-BY-SA 4.0
• Original sample rate: 32 kHz (resampled to 16 kHz here)
• Citation: Sanabria et al. (2023), 'The Edinburgh International Accents of English Corpus', ICASSP 2023.
• Note: EdAcc ships only validation + test splits (it is designed as an evaluation benchmark). Repurposing it as fine-tuning data is fine, but if you also EVALUATE on EdAcc, hold its speakers out — the split_src field is preserved so you can.

DailyTalk (source=dailytalk, 23292 clips)

• HuggingFace: eustlb/dailytalk (audio mirror of the original DailyTalk)
• License: CC-BY-SA 4.0
• Original sample rate: 24 kHz (resampled to 16 kHz here)
• Citation: Lee, Park, Kim (2023), 'DailyTalk: Spoken Dialogue Dataset for Conversational Text-to-Speech', ICASSP 2023.
• Note: Scripted speech: no genuine disfluency / overlap. Use to reinforce everyday-life register, not to model spontaneity. Only 2 voices.