RobotsMali/an-be-kalan-bench

Bambara Educational Speech Dataset

This dataset is a collection of READ Bambara text based on educational children's books from RobotsMali's GAIFE project. It is designed to support the training and benchmarking of Automatic Speech Recognition (ASR) models, with a particular focus on child speech, regional acoustics, and repetitive text structures (inherent to the domain).

The dataset is structured into two separate subsets to support specialized training and evaluation paradigms:

1. main: Contains separate training and test splits with disjoint recorded book titles and speaker IDs. Speaker IDs may not reliably identify distinct people across sessions.
2. duplicate: A highly dense, multi-speaker redundant training set featuring multiple recordings of the same source literature by a diverse pool of speakers.

---

Dataset Architecture & Splits

The release has main/train, main/test, and duplicate/train. Exact BookTitle strings do not overlap between the test and either training split. The two training splits contain the same 22 title strings. Speaker IDs are metadata identifiers rather than verified distinct people; the separate main and duplicate training sets share IDs, while the test has no ID overlap with either training split at this revision. Dataset-derived totals below use the published duration metadata and count each utterance, including repeated readings.

Summary Statistics

Metric	main/train	main/test	duplicate/train	All published rows
Utterances	1,203	724	33,481	35,408
Total duration	1.624 h	0.889 h	44.018 h	46.531 h
Distinct speaker IDs	8	11	113	124
Distinct book titles	22	17	22	39
Mean audio duration	4.86 s	4.42 s	4.73 s	4.73 s
Audio duration range	0.72–25.60 s	0.32–16.48 s	0.08–37.52 s	0.08–37.52 s
Mean sentence length	8.06 words	7.18 words	7.85 words	7.84 words
Sentence length range	1–26 words	1–21 words	1–26 words	1–26 words
Rows missing any age, gender, or speaker ID	0	38	0	38

The two training splits total 34,684 utterances and 45.642 h (main/train: 1.624 h; duplicate/train: 44.018 h). All published rows, including the 0.889 h test split, sum to 46.531 h. These are retained dataset durations, not the 55 hours of raw campaign recordings.

Demographic distributions

Known speakerAge values span 5–19 in this release. The test split has 2 utterances labelled age 5; these are metadata values that merit source-record verification. The test also has 38 rows with unknown age and gender. The age bands below count utterances, not unique children; 10–15 and 16–20 include both endpoints.

Age band	main/train	main/test	duplicate/train	All published rows
Under 10	67	93	1,782	1,942
10–15	1,097	527	17,459	19,083
16–20	39	66	14,240	14,345
Unknown	0	38	0	38

Gender metadata	main/train	main/test	duplicate/train	All published rows
Female	783	522	16,688	17,993
Male	420	164	16,793	17,377
Unknown	0	38	0	38

At the current release revision, the test duration field places 692/724 utterances (95.6%) at or below 10 s and 719/724 (99.3%) at or below 15 s. These are metadata filter counts, not proof of a historical validation run's exact row count.

Book inventory and age distribution

The table uses the exact published BookTitle strings. Ages are utterance counts in the order under 10 / 10–15 / 16–20 / unknown. Speaker IDs are distinct recorded identifiers within a title, not verified readers or complete-read counts.

main/train — 22 titles

BookTitle	Utterances	Duration (h)	Speaker IDs	Age counts (<10 / 10–15 / 16–20 / unknown)
Aminata Fari Fisayara	68	0.089	2	67 / 1 / 0 / 0
Bakɔrɔnin Saba	45	0.057	1	0 / 45 / 0 / 0
Bɛnkɛ Tɔm Ka So	124	0.182	2	0 / 124 / 0 / 0
Dawuda ni a Mɔkɛ	38	0.054	1	0 / 38 / 0 / 0
Dɔgɔtɔrɔ ni Farafinfurabɔla a	97	0.144	1	0 / 97 / 0 / 0
Filomani	65	0.089	1	0 / 65 / 0 / 0
Gawusu ni Masakɛ Sidiki	51	0.071	1	0 / 51 / 0 / 0
Gerenadi-Feerew	47	0.081	1	0 / 47 / 0 / 0
Gesedala Musa	35	0.045	1	0 / 35 / 0 / 0
Gundola Kuma	59	0.064	2	0 / 59 / 0 / 0
Kalo la Taama	64	0.077	1	0 / 64 / 0 / 0
Kan Orobotik	39	0.046	1	0 / 39 / 0 / 0
Korokara Yɛrɛdɔnbali	74	0.099	2	0 / 74 / 0 / 0
Kurun	52	0.059	1	0 / 52 / 0 / 0
Lamini Ka Don Kɛrɛnkɛrɛnnen	77	0.099	1	0 / 77 / 0 / 0
Mama ka Sama	33	0.051	1	0 / 33 / 0 / 0
Ne ni Mama ka Gafe Kalan	38	0.037	1	0 / 38 / 0 / 0
Saratu	55	0.101	1	0 / 55 / 0 / 0
Subahana Daga	39	0.064	1	0 / 0 / 39 / 0
Sɔminiminɛnw	28	0.023	1	0 / 28 / 0 / 0
Tulonkɛw	51	0.057	2	0 / 51 / 0 / 0
Yɛlɛ Ka Di Npogotiginin Mi Ye	24	0.036	1	0 / 24 / 0 / 0

main/test — 17 titles

BookTitle	Utterances	Duration (h)	Speaker IDs	Age counts (<10 / 10–15 / 16–20 / unknown)
Anw Bɛ Baara Kɛ!	55	0.106	3	50 / 5 / 0 / 0
Ayisa ye nkalontigɛ dabila	55	0.076	2	0 / 54 / 1 / 0
Bako Cɛnin Ŋaniya Ɲuman	46	0.079	1	0 / 46 / 0 / 0
Bama Miirina	24	0.024	2	22 / 0 / 1 / 1
Cɛni Tulogɛlɛn	46	0.057	1	0 / 9 / 0 / 37
Denmisɛnya Kojuguw	37	0.029	2	0 / 37 / 0 / 0
Donfɛnw	27	0.026	1	0 / 27 / 0 / 0
Fali Nalonma Ni Ba Kegunma ani Ɲininkaliw ni u j	39	0.057	2	0 / 39 / 0 / 0
Jate	36	0.039	1	0 / 0 / 36 / 0
Ji Poyi Yɔrɔ	72	0.058	1	0 / 72 / 0 / 0
Kewale Numanw	26	0.032	1	0 / 26 / 0 / 0
Kogo	39	0.075	1	0 / 39 / 0 / 0
Kulɔriw	30	0.033	1	0 / 30 / 0 / 0
Mali kunkanko	92	0.083	1	0 / 92 / 0 / 0
Namasatigi	21	0.027	1	21 / 0 / 0 / 0
Ne ni Papa ka Gafe Kalan	42	0.037	3	0 / 14 / 28 / 0
Ni a tun bɛ se ...	37	0.051	1	0 / 37 / 0 / 0

duplicate/train — 22 titles

BookTitle	Utterances	Duration (h)	Speaker IDs	Age counts (<10 / 10–15 / 16–20 / unknown)
Aminata Fari Fisayara	2,420	2.792	38	172 / 1230 / 1018 / 0
Bakɔrɔnin Saba	1,596	2.131	38	126 / 790 / 680 / 0
Bɛnkɛ Tɔm Ka So	3,212	3.687	28	111 / 1503 / 1598 / 0
Dawuda ni a Mɔkɛ	1,179	1.585	31	40 / 590 / 549 / 0
Dɔgɔtɔrɔ ni Farafinfurabɔla a	2,106	2.868	24	194 / 850 / 1062 / 0
Filomani	2,124	2.552	32	134 / 1050 / 940 / 0
Gawusu ni Masakɛ Sidiki	1,473	2.022	30	148 / 856 / 469 / 0
Gerenadi-Feerew	1,330	2.100	29	48 / 753 / 529 / 0
Gesedala Musa	966	1.349	27	72 / 536 / 358 / 0
Gundola Kuma	1,710	2.257	28	177 / 941 / 592 / 0
Kalo la Taama	1,731	2.111	27	67 / 994 / 670 / 0
Kan Orobotik	1,200	1.833	30	83 / 581 / 536 / 0
Korokara Yɛrɛdɔnbali	1,873	2.357	25	0 / 1094 / 779 / 0
Kurun	1,314	1.859	27	0 / 736 / 578 / 0
Lamini Ka Don Kɛrɛnkɛrɛnnen	1,836	2.491	25	26 / 985 / 825 / 0
Mama ka Sama	826	1.415	25	0 / 442 / 384 / 0
Ne ni Mama ka Gafe Kalan	1,167	1.367	27	122 / 722 / 323 / 0
Saratu	1,360	2.353	26	58 / 672 / 630 / 0
Subahana Daga	1,198	1.580	29	84 / 652 / 462 / 0
Sɔminiminɛnw	1,236	1.160	32	120 / 564 / 552 / 0
Tulonkɛw	1,094	1.285	20	0 / 605 / 489 / 0
Yɛlɛ Ka Di Npogotiginin Mi Ye	530	0.867	20	0 / 313 / 217 / 0

Statistics above were computed from all metadata rows in the dataset's Parquet shards at revision 95cf3103994ac13c13e086518c320a614cf75085. Sentence length counts whitespace-separated words; duration uses the published duration field. Audio was not decoded for this metadata audit.

---

Critical Considerations: Features vs. Weaknesses

When developing models on this dataset, users should balance its unique profile against known recording constraints:

1. The High-Volume Duplication Matrix

• As a Feature: Due to the severe scarcity of open-source text and educational literature in Bambara, collecting deep audio variations on a finite set of text was a deliberate design choice. This split provides a robust playground for specific speech experiments, such as acoustic multi-speaker verification, text-constrained acoustic profiling, and downstream speech representation probing.
• As a Weakness: The textual diversity in the duplicate subset is inherently bottlenecked by the underlying literature. Models trained aggressively on the duplicate split without constraint can rapidly overfit to the vocabulary, tone structures, and phonetic bounds of these 22 recorded book titles.

2. Known Metadata Inconsistencies

• The Identifier Inconsistency: The release has 113 distinct speaker IDs across its training splits and 124 across all splits. On-the-ground project coordinators reported approximately 60 actual speakers; that estimate has not been verified from these metadata rows.
• Implication: This discrepancy highlights an operational metadata inflation error where individual speakers were assigned differing tracking IDs across different recording sessions, days, or environments. Users should exercise caution when benchmarking strict zero-shot speaker verification algorithms on this dataset without manual speaker clustering.

---

Dataset Format

The Hugging Face release uses Parquet shards with BookTitle, sentenceID, text, speakerAge, speakerGender, speakerID, duration, and an audio feature. A local export can derive a NeMo-compatible JSON Lines (.jsonl) manifest. In that derived format, audio_filepath is a local file path, not a field of the published Parquet rows. For example:

{
  "BookTitle": "Aminata Fari Fisayara",
  "sentenceID": 1,
  "text": "Mɔgɔ caman tun bɛ yen.",
  "speakerAge": 15,
  "speakerGender": "male",
  "speakerID": "spk_1778182779107_12bb3ea9c8",
  "duration": 2.56,
  "audio_filepath": "data/audios/dup_Aminata_Fari_Fisayara_1_1200.wav"
}

Citation

If you utilize this dataset or its subsets in research, please cite the repository card details accordingly.

BIBTEX ENTRY COMING SOON