Document the ASR v2 splits in the main dataset card

Moves the v2 documentation into this card, where readers will find it,
and drops the separate data/ASR_v2/README.md, which nothing rendered.

Two insertions only: a Table of Contents entry and a section at the end of
Data Splits. The front matter and every other line are byte-identical.

Condensed to corpus-level figures and the three things that actually trip
people up when reading v2; per-language detail stays in
data/ASR_v2/metadata/split_comparison.csv.

README.md

@@ -925,6 +925,7 @@ The WAXAL dataset is a large-scale multilingual speech corpus for African langua
     -   [ASR Data Fields](#asr-data-fields)
     -   [TTS Data Fields](#tts-data-fields)
     -   [Data Splits](#data-splits)
+    -   [ASR v2 splits](#asr-v2-splits-speaker-disjoint)
 -   [Dataset Curation](#dataset-curation)
 -   [Considerations for Using the Data](#considerations-for-using-the-data)
 -   [Additional Information](#additional-information)
@@ -1097,6 +1098,61 @@ transcription.
 The **TTS Dataset** follows a similar structure, with data split into `train`,
 `validation`, and `test` sets.
 
+### ASR v2 splits (speaker-disjoint)
+
+The ASR half also ships a second set of splits intended for benchmarking, under
+`data/ASR_v2/`. They re-partition the same utterances so that **no speaker
+appears in more than one split**, meaning a model is always evaluated on voices
+it did not train on. The audio is untouched -- v2 is a re-labelling, not a new
+release of the recordings.
+
+| Across all 19 ASR languages | original | v2 |
+|---|---|---|
+| Eval speakers also present in train | up to 100 % | **0 %** |
+| Eval utterances with an exact transcript twin in train | 1.09 % | **0.09 %** |
+| Split sizes by duration | 80 / 10 / 10 nominal | 84.5 / 7.7 / 7.7 |
+
+438,230 utterances, 2,242 hours. Gender is matched across splits for the
+12 languages that carry gender labels. Per-language figures are in
+`data/ASR_v2/metadata/split_comparison.csv`, and the full provenance -- seed,
+source revision, every parameter -- in `metadata/splits_manifest.json`.
+
+#### Reading v2
+
+Read the existing ASR configs and re-label each row from the split map:
+
+```python
+import pandas as pd
+from datasets import load_dataset
+
+LANG, WANT = "sna", "test"
+m = pd.read_csv(
+    f"hf://datasets/google/WaxalNLP/data/ASR_v2/metadata/split_map/{LANG}.csv",
+    dtype=str,
+)
+v2 = dict(zip(m["id"], m["v2_split"]))       # `id` is unique within a language
+
+for v1_split in ("train", "validation", "test"):          # all three
+    for row in load_dataset("google/WaxalNLP", f"{LANG}_asr",
+                            split=v1_split, streaming=True):
+        if v2.get(row["id"]) == WANT:
+            ...
+```
+
+Three things worth knowing:
+
+- **Read all three original labelled splits.** Utterances move between them:
+  Acholi's v2 test set draws 397 of its 515 rows from the original *train*.
+  Reading only the same-named split returns a fraction of the data, silently.
+- **Skip `unlabeled`.** It carries no transcripts and takes no part in the
+  re-split.
+- **Two utterances are unmapped** -- `kpo_149159` has no `speaker_id` and
+  `lin_9193` has an empty transcript. They are listed in
+  `metadata/excluded_rows.csv`; `v2.get()` returns `None` for them.
+
+Streaming bounds memory rather than download: filtering still transfers the
+language's labelled half, 1.6 GB (`ach`) to 12.7 GB (`sid`).
+
 ## Dataset Curation
 
 The data was gathered by multiple partners:

data/ASR_v2/README.md

@@ -1,389 +0,0 @@
----
-language:
-- ach
-- aka
-- amh
-- dag
-- dga
-- ewe
-- ful
-- kpo
-- lin
-- lug
-- mlg
-- myx
-- nyn
-- orm
-- sid
-- sna
-- tir
-- wal
-- xog
-license:
-- CC-BY-4.0
-- CC-BY-SA-4.0
-task_categories:
-- automatic-speech-recognition
-source_datasets:
-- google/WaxalNLP
-multilinguality:
-- multilingual
-annotation_creators:
-- human-annotated
-size_categories:
-- 100K<n<1M
-pretty_name: WAXAL ASR — speaker-disjoint re-split (v2)
-tags:
-- audio
-- automatic-speech-recognition
-- african-languages
-- speaker-disjoint-splits
-- resplit
----
-
-# WAXAL ASR — speaker-disjoint re-split (v2)
-
-A speaker-disjoint re-split of the ASR half of this dataset, published by the
-dataset maintainers alongside v1. **v1 is unchanged and stays supported** --
-every existing config resolves to exactly the files it always did.
-
-**The audio is unchanged.** Every payload is the byte-for-byte MP3 from v1;
-this is a re-partitioning of the same recordings, not a re-encode and not a
-quality filter. What changes is *which utterances are in which split*.
-
-- 438,230 labelled utterances, 2,242 hours,
-  19 languages
-- v1 remains untouched and fully usable: `load_dataset("google/WaxalNLP", "ach_asr")`
-- **The repacked audio is not published yet.** What is here is the split
-  definition: every utterance's v1 and v2 split, for all 19 languages. It is
-  complete and final -- the audio is a materialisation of it, not a
-  precondition for using it. Apply it to your existing copy of v1 with the
-  snippet below.
-
-## Why
-
-The v1 ASR splits are not safe to benchmark on. Measured across all
-19 languages:
-
-| Code | Language | v1 eval speakers also in train | v2 | v1 eval transcripts also in train | v2 | v2 leakage vs chance |
-|---|---|---|---|---|---|---|
-| `ach` | Acholi | 99.0 % | **0.0 %** | 0.6 % | 1.5 % | 0.88 (-2.8σ) |
-| `aka` | Akan | 100.0 % | **0.0 %** | 0.0 % | 0.0 % | 0.95 (-0.9σ) |
-| `amh` | Amharic | 0.0 % | **0.0 %** | 0.1 % | 0.4 % | 0.70 (-6.0σ) |
-| `dag` | Dagbani | 97.9 % | **0.0 %** | 0.2 % | 0.1 % | 0.75 (-9.3σ) |
-| `dga` | Dagaare | 100.0 % | **0.0 %** | 19.1 % | 0.2 % | 0.83 (-3.0σ) |
-| `ewe` | Ewe | 99.8 % | **0.0 %** | 0.0 % | 0.0 % | 0.86 (-3.5σ) |
-| `ful` | Fula | 99.4 % | **0.0 %** | 0.0 % | 0.0 % | 0.97 (-0.5σ) |
-| `kpo` | Ikposo | 99.1 % | **0.0 %** | 0.0 % | 0.1 % | 0.76 (-4.4σ) |
-| `lin` | Lingala | 98.7 % | **0.0 %** | 0.1 % | 0.1 % | 1.06 (+1.0σ) |
-| `lug` | Luganda | 99.6 % | **0.0 %** | 0.3 % | 0.3 % | 0.82 (-5.8σ) |
-| `mas` | Masaaba | 98.4 % | **0.0 %** | 0.2 % | 0.2 % | 0.80 (-4.8σ) |
-| `mlg` | Malagasy | 97.7 % | **0.0 %** | 0.0 % | 0.0 % | 0.90 (-1.4σ) |
-| `nyn` | Runyankole | 98.5 % | **0.0 %** | 0.7 % | 0.5 % | 0.78 (-6.1σ) |
-| `orm` | Oromo | 0.0 % | **0.0 %** | 0.0 % | 0.2 % | 0.51 (-7.9σ) |
-| `sid` | Sidama | 0.0 % | **0.0 %** | 0.0 % | 0.2 % | 0.67 (-4.8σ) |
-| `sna` | Shona | 97.0 % | **0.0 %** | 0.1 % | 0.1 % | 1.09 (+1.0σ) |
-| `sog` | Soga | 98.3 % | **0.0 %** | 0.0 % | 0.0 % | 0.87 (-3.8σ) |
-| `tir` | Tigrinya | 0.0 % | **0.0 %** | 2.0 % | 0.1 % | 0.69 (-4.7σ) |
-| `wal` | Wolaytta | 0.0 % | **0.0 %** | 0.0 % | 0.1 % | 0.83 (-3.5σ) |
-
-A model that has heard the test speakers during training reports a score that
-is partly speaker memorisation rather than recognition, and the size of that
-effect is not knowable after the fact. The re-split removes the channel rather
-than trying to correct for it.
-
-## What v2 guarantees
-
-Every figure here was measured from the published split maps joined to v1, not
-asserted from the code that produced them.
-
-| Property | v1 | v2 |
-|---|---|---|
-| Languages whose eval speakers also appear in train | 14 of 19 | **0 of 19** |
-| Worst eval speaker leakage, any language | 100.0 % | **0.0 %** |
-| Corpus-wide eval utterances with an exact twin in train | 1.09 % | **0.094 %** |
-| Leakiest language in v1, `dga` | 19.1 % | **0.2 %** |
-| Leakiest language in v2, `ach` | 0.6 % | 1.5 % |
-| Languages with gender labels | 12 of 19 | 12 of 19 |
-
-**Speaker-disjointness is structural, not statistical.** The unit of assignment
-is the speaker, so an utterance cannot cross a split boundary without its
-speaker. It is not a threshold that was met; it is a property the construction
-cannot violate.
-
-**Sizes are 80/10/10 by duration, with eval splits targeted at ~12 hours.**
-That target binds on the six largest languages, which therefore evaluate on
-roughly 5.4 % per side rather than 10 %; the surplus goes to train. It is a
-target, not a hard ceiling -- five eval splits land between 12.0 and 12.2 h,
-and the nominal 2,000-utterance bound carries 35 % slack, so 14 of 38 eval
-splits exceed 2,000 utterances (largest: `tir` validation at 2,686). Corpus-wide this comes to
-84.5 %/7.7 %/7.7 % by duration. Per-language figures are in the table below.
-
-**What v2 does not change.** No quality filtering, no re-encoding, no
-resampling. The same utterances and the same audio as v1, re-partitioned.
-
-## How the splits were built
-
-Speakers are the unit of assignment, so **speaker-disjointness is structural**:
-an utterance cannot cross a split boundary without its speaker, and a speaker
-belongs to exactly one split. The assignment is then chosen to satisfy, in
-order of precedence:
-
-1. **Speaker-disjointness** — a hard constraint, enforced by construction.
-2. **Low train/eval lexical leakage** — a `sqrt(idf)`-weighted affinity over
-   shared sentences, word trigrams and word types is minimised across the
-   train/eval boundary. Because inverse document frequency goes to zero for
-   items nearly everyone uses, *shared common vocabulary is invisible to the
-   objective* and only shared rare content is penalised. This is deliberate:
-   common-word overlap between train and test is coverage, not leakage, and a
-   test set that avoided it would measure a distribution nobody deploys against.
-3. **Gender balance** — the female share of each eval split is matched to the
-   corpus, weighted by reference words (the weighting under which a
-   micro-averaged WER is or is not gender-representative).
-
-Targets are 80/10/10 by duration, but two bounds move that in practice, and
-both are recorded per language in `splits_manifest.json`: the eval share is
-allowed to **grow** to reach a 20,000-reference-word floor in the smallest
-languages, and is **capped** at 12 hours or 2,000 utterances per eval split in
-the largest — so a 220-hour language like Amharic evaluates on about 5.5 % per
-side rather than 10 %, with the surplus going to train. No single speaker may
-exceed 25 % of an eval split. The search is a randomised
-longest-processing-time greedy seed followed by steepest-descent local search,
-seeded from `20260730` and fully deterministic; re-running reproduces
-the assignment hash recorded per language in `splits_manifest.json`.
-
-Every goal is scaled against a **null distribution** of
-200 random size-matched speaker-disjoint splits
-of the same language, so the reported numbers can be read against chance.
-
-## Splits
-
-Train / validation / test, as utterances, hours and distinct speakers:
-
-| Code | train | validation | test |
-|---|---|---|---|
-| `ach` | 4,120 / 26.0 h / 263 | 520 / 3.2 h / 33 | 515 / 3.2 h / 33 |
-| `aka` | 10,196 / 55.6 h / 110 | 1,287 / 7.0 h / 11 | 1,269 / 6.9 h / 13 |
-| `amh` | 39,455 / 196.0 h / 550 | 2,473 / 12.0 h / 34 | 2,426 / 12.0 h / 34 |
-| `dag` | 14,237 / 77.2 h / 884 | 1,798 / 9.7 h / 79 | 1,784 / 9.7 h / 107 |
-| `dga` | 15,091 / 83.8 h / 275 | 1,887 / 10.5 h / 36 | 1,896 / 10.5 h / 38 |
-| `ewe` | 15,092 / 79.8 h / 431 | 1,885 / 10.0 h / 53 | 1,884 / 10.0 h / 54 |
-| `ful` | 19,293 / 100.6 h / 168 | 2,303 / 12.0 h / 21 | 2,285 / 11.9 h / 22 |
-| `kpo` | 14,416 / 82.3 h / 369 | 1,803 / 10.3 h / 46 | 1,800 / 10.3 h / 46 |
-| `lin` | 14,523 / 72.3 h / 74 | 1,778 / 9.1 h / 13 | 1,808 / 9.1 h / 10 |
-| `lug` | 5,366 / 36.7 h / 270 | 692 / 4.7 h / 35 | 699 / 4.7 h / 35 |
-| `mas` | 6,856 / 39.2 h / 290 | 857 / 4.9 h / 36 | 861 / 4.9 h / 36 |
-| `mlg` | 18,467 / 94.5 h / 199 | 2,316 / 11.8 h / 25 | 2,303 / 11.8 h / 27 |
-| `nyn` | 6,752 / 40.8 h / 293 | 861 / 5.1 h / 37 | 859 / 5.2 h / 37 |
-| `orm` | 40,141 / 198.7 h / 361 | 2,473 / 11.9 h / 23 | 2,431 / 12.0 h / 22 |
-| `sid` | 40,547 / 202.0 h / 463 | 2,444 / 11.9 h / 28 | 2,413 / 12.0 h / 28 |
-| `sna` | 14,079 / 79.5 h / 131 | 1,744 / 9.9 h / 20 | 1,762 / 9.9 h / 17 |
-| `sog` | 6,298 / 40.4 h / 260 | 786 / 5.1 h / 33 | 793 / 5.0 h / 32 |
-| `tir` | 44,945 / 194.8 h / 408 | 2,686 / 12.1 h / 25 | 2,684 / 12.1 h / 24 |
-| `wal` | 42,258 / 195.1 h / 381 | 2,572 / 12.2 h / 23 | 2,461 / 12.0 h / 23 |
-
-### Gender balance
-
-Female share of reference words. `n/a` means the source has **no gender labels
-at all** for that language — not that balance was achieved.
-
-| Code | train | validation | test |
-|---|---|---|---|
-| `ach` | 0.268 | 0.267 | 0.264 |
-| `aka` | n/a | n/a | n/a |
-| `amh` | 0.454 | 0.454 | 0.454 |
-| `dag` | n/a | n/a | n/a |
-| `dga` | n/a | n/a | n/a |
-| `ewe` | n/a | n/a | n/a |
-| `ful` | n/a | n/a | n/a |
-| `kpo` | n/a | n/a | n/a |
-| `lin` | 0.550 | 0.548 | 0.539 |
-| `lug` | 0.470 | 0.471 | 0.470 |
-| `mas` | 0.306 | 0.307 | 0.306 |
-| `mlg` | n/a | n/a | n/a |
-| `nyn` | 0.452 | 0.454 | 0.452 |
-| `orm` | 0.562 | 0.561 | 0.559 |
-| `sid` | 0.654 | 0.656 | 0.660 |
-| `sna` | 0.614 | 0.612 | 0.600 |
-| `sog` | 0.479 | 0.477 | 0.482 |
-| `tir` | 0.552 | 0.556 | 0.552 |
-| `wal` | 0.661 | 0.661 | 0.656 |
-
-### Languages needing attention
-
-| Code | Status | Note |
-|---|---|---|
-| `ach` | ok | test: 19,399 reference words against the 20,000 target (WER standard error ~0.36 pp before clustering) |
-| `amh` | ok | eval share reduced 0.100 -> 0.055 by the 12-hour eval cap |
-| `ful` | ok | eval share reduced 0.100 -> 0.096 by the 12-hour eval cap |
-| `lug` | ok | eval share raised 0.100 -> 0.102 to reach the 20,000-reference-word floor; validation: 18,292 reference words against the 20,000 target (WER standard error ~0.37 pp before clustering); test: 18,928 reference words against the 20,000 target (WER standard error ~0.36 pp before clustering) |
-| `mas` | ok | validation: 19,731 reference words against the 20,000 target (WER standard error ~0.35 pp before clustering); test: 14,265 reference words against the 20,000 target (WER standard error ~0.41 pp before clustering) |
-| `nyn` | ok | eval share raised 0.100 -> 0.101 to reach the 20,000-reference-word floor; validation: 17,439 reference words against the 20,000 target (WER standard error ~0.37 pp before clustering); test: 19,041 reference words against the 20,000 target (WER standard error ~0.36 pp before clustering) |
-| `orm` | ok | eval share reduced 0.100 -> 0.054 by the 12-hour eval cap |
-| `sid` | ok | eval share reduced 0.100 -> 0.053 by the 12-hour eval cap |
-| `tir` | ok | eval share reduced 0.100 -> 0.055 by the 12-hour eval cap |
-| `wal` | ok | eval share reduced 0.100 -> 0.055 by the 12-hour eval cap |
-
-## Reading v2 today
-
-The audio is not republished. v2 re-partitions v1's rows, so you reconstruct a
-v2 split by loading v1 and re-labelling each row from the split map. Nothing is
-added or removed: the same 438,230 utterances, re-assigned.
-
-```python
-import pandas as pd
-from datasets import load_dataset
-
-LANG, WANT = "sna", "test"
-
-m = pd.read_csv(
-    f"hf://datasets/google/WaxalNLP/data/ASR_v2/metadata/split_map/{LANG}.csv"
-)
-v2 = dict(zip(m["id"], m["v2_split"]))          # `id` is unique within a language
-
-for v1_split in ("train", "validation", "test"):        # all three -- see below
-    ds = load_dataset("google/WaxalNLP", f"{LANG}_asr",
-                      split=v1_split, streaming=True)
-    for row in ds:
-        if v2.get(row["id"]) == WANT:
-            ...                                  # a v2 `{WANT}` utterance
-```
-
-`pd.read_csv` on an `hf://` path needs `huggingface_hub` installed (it is, if
-`datasets` is). Pass `dtype=str` -- some ids begin with an underscore.
-
-**Budget the transfer.** Streaming does not save bandwidth here: there is no
-predicate pushdown and the audio bytes are inline in the same row groups, so
-filtering still pulls the language's whole labelled half. Measured: Shona's v2
-test set is 1,749 rows but costs 552 MB and about 8 minutes. Reconstructing all
-three splits of one language ranges from 1.6 GB (`ach`) to 12.7 GB (`sid`).
-Streaming's advantage is bounded memory, not bounded download.
-
-### Two mistakes that fail silently
-
-**Read all three v1 labelled splits.** Utterances move between them. Acholi's
-v2 `test` set of 515 rows is drawn 397 from v1 *train*, 58 from v1 validation
-and 60 from v1 test. Reading only v1's `test` split returns **60 rows instead
-of 515** -- 88 % of your evaluation set silently missing, with no error. (It
-does *not* reintroduce speaker leakage: v2 moves whole speakers, so any subset
-of one v2 split stays disjoint from the others. The damage is that you are
-measuring on a twelfth of the data, and on a non-random twelfth.)
-
-**Never read the `unlabeled` split.** It carries no transcripts, takes no part
-in the re-split, and is by far the largest thing in the repository (637 GB
-across the 19 languages). `v2.get(row["id"])` returns `None` for all of it, so
-the filter silently yields nothing after a very long wait.
-
-### 2 utterances are deliberately unmapped
-
-`v2.get(...)` returns `None` for 2 of v1's labelled rows, which are excluded
-from the splits rather than assigned: `kpo_149159` has an empty `speaker_id`
-(so it cannot be placed without risking leakage in one direction or the other)
-and `lin_9193`'s transcript is a bare newline. Both are listed in
-`metadata/excluded_rows.csv`. Treat `None` as "not in v2", not as an error.
-
-
-## Schema
-
-**What you read today** is v1's schema — `id`, `speaker_id`, `transcription`,
-`language`, `gender`, `audio` — with the v2 split supplied by the split map:
-
-| Column | Notes |
-|---|---|
-| `language`, `id` | join key; `id` is unique within a language |
-| `v1_split` | which v1 split the utterance is stored in |
-| `v2_split` | `train` / `validation` / `test` |
-
-**Note `speaker_id` is only unique within a language.** 354 speaker ids are
-reused across different languages, so grouping by `speaker_id` alone over a
-multilingual concatenation merges unrelated speakers. Group by
-`(language, speaker_id)`.
-
-<details>
-<summary>Schema of the repacked files, once the audio is published</summary>
-
-The six v1 columns, unchanged, plus derived fields:
-
-| Column | Notes |
-|---|---|
-| `id`, `speaker_id`, `transcription`, `language`, `gender` | unchanged from v1 |
-| `audio` | `struct<bytes, path>` — **byte-identical** to v1 |
-| `v1_split` | which v1 split this utterance came from |
-| `duration_s`, `n_bytes` | derived; the corpus is 128 kbit/s CBR MP3 |
-| `sample_rate`, `channels` | measured **per row**; heterogeneous (see limitations) |
-| `provider`, `licence` | the licence travels with the row |
-
-The stray `__index_level_0__` column present in 36 of the v1 shards is dropped,
-and every v2 file uses 100-row row groups.
-
-</details>
-
-## Limitations
-
-- **No gender labels at all** for `aka`, `dag`, `dga`, `ewe`, `ful`, `kpo`, `mlg`.
-  The gender goal is undefined there, not satisfied, and is reported as `n/a`.
-- **Heterogeneous audio format**, and it varies *within* a shard, not just
-  between languages: 16 kHz mono (`ach`, `lug`), 44.1 kHz stereo (`aka`, `dag`,
-  `dga`, `ewe`), 48 kHz mono (`ful`, `lin`, `sna`, `tir`, `wal`, most `amh`)
-  and 48 kHz stereo (some `amh`). One `amh` test row group holds 14 stereo and
-  6 mono clips. v1 documents none of this. Nothing is resampled -- a silent
-  normalisation cannot be verified against the original.
-- **This is a re-split, not a clean-up.** No quality filtering has been
-  applied. Utterance-level defects present in v1 are present here.
-- **The `unlabeled` split is out of scope** (637 GB, no transcripts). Take it
-  from the v1 config, which is unchanged.
-- **`speaker_id` is trusted, not verified acoustically** unless the manifest
-  says otherwise. If an id turns out to cover more than one voice, the
-  disjointness guarantee is over ids, not over voices.
-- The `v1 eval transcripts also in train` column is an **exact census** over
-  every eval transcript. A separate fuzzy near-duplicate rate (0.95 similarity)
-  is recorded in `metadata/split_comparison.csv` and is computed on a sample of
-  at most 1,200 eval transcripts per split, with the sample
-  size stored beside it.
-- Rows with an empty or placeholder transcript, or with no `speaker_id`, are
-  excluded from the splits and listed in `metadata/excluded_rows.csv`.
-  Nothing else is filtered.
-
-## Licensing
-
-The corpus mixes two licences by contributing partner. **Each language is a
-separate config specifically so that aggregating them does not relicense the
-CC-BY-4.0 languages under ShareAlike.** The `licence` column travels with every
-row.
-
-| Provider | Languages | Licence |
-|---|---|---|
-| Digital Umuganda | Amharic (`amh`), Fula (`ful`), Lingala (`lin`), Malagasy (`mlg`), Oromo (`orm`), Shona (`sna`), Sidama (`sid`), Tigrinya (`tir`), Wolaytta (`wal`) | `CC-BY-SA-4.0` |
-| Makerere University | Acholi (`ach`), Luganda (`lug`), Masaaba (`mas`), Runyankole (`nyn`), Soga (`sog`) | `CC-BY-SA-4.0` |
-| University of Ghana | Akan (`aka`), Dagaare (`dga`), Dagbani (`dag`), Ewe (`ewe`), Ikposo (`kpo`) | `CC-BY-4.0` |
-
-Derivatives of the CC-BY-SA-4.0 languages must carry the same terms. Combining
-configs means inheriting ShareAlike for the combined work.
-
-## Attribution and citation
-
-Source corpus: [`google/WaxalNLP`](https://huggingface.co/datasets/google/WaxalNLP),
-revision `e0a62aaebc61`, collected by Makerere
-University, the University of Ghana and Digital Umuganda, funded by Google and
-the Gates Foundation. Upstream sources:
-[UGSpeechData](https://doi.org/10.57760/sciencedb.22298),
-[AfriVoice](https://huggingface.co/datasets/DigitalUmuganda/AfriVoice), and the
-[Yogera Dataset](https://doi.org/10.7910/DVN/BEROE0).
-
-```bibtex
-@article{waxal2026,
-  title={WAXAL: A Large-Scale Multilingual African Language Speech Corpus},
-  author={Anonymous},
-  journal={arXiv preprint arXiv:2602.02734},
-  year={2026}
-}
-```
-
-## Reproducing this
-
-The split is fully determined by the manifest. `splits_manifest.json` records
-the seed, the pinned source revision, every tolerance, the library versions,
-and a hash of the text-canonicalisation function that all leakage figures were
-computed through. `metadata/split_map/{lang}.csv` carries the per-utterance
-assignment, and `metadata/split_comparison.csv` the full v1-versus-v2 audit
-table these figures are drawn from.