Professor commited on
Commit
41833bb
·
verified ·
1 Parent(s): 94ebd9c

Add dataset card

Browse files
Files changed (1) hide show
  1. README.md +95 -0
README.md ADDED
@@ -0,0 +1,95 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language: [lg]
3
+ license: cc0-1.0
4
+ multilinguality: monolingual
5
+ task_categories: [text-to-speech, automatic-speech-recognition]
6
+ tags: [luganda, ganda, speech, tts, asr, african-languages, common-voice]
7
+ pretty_name: Luganda Speech Data (Pooled)
8
+ size_categories: [10K<n<100K]
9
+ ---
10
+
11
+ # Luganda Speech Data (Pooled)
12
+
13
+ A **~111.4-hour** Luganda speech corpus, drawn from Mozilla Common Voice and
14
+ filtered to the `train` split. Part of the
15
+ [AfroNet](https://github.com/osinkolu/afronet-tts-data) multi-language TTS data
16
+ effort.
17
+
18
+ ## Source
19
+
20
+ **Mozilla Common Voice 26.0**, Luganda (`lg`) locale, accessed via the
21
+ [Mozilla Data Collective](https://mozilladatacollective.com/datasets/cmqinhruw00vinq07tohj4fd3)
22
+ mirror. 67,269 clips, 111.4h, source `dataset_id`/`source` = `common_voice`.
23
+
24
+ Common Voice is crowdsourced, scripted read speech: contributors record themselves
25
+ reading prompted sentences, and other contributors validate the recordings. The full
26
+ corpus splits into `train`/`dev`/`test`/`validated` (a superset of all three plus
27
+ extra validated-but-unsplit clips) -- only **`train`** is used here, same
28
+ held-out-partition policy we apply to DSN African Voices' splits.
29
+
30
+ **Why not WAXAL's `lug_tts` config too?** WAXAL's Luganda TTS config (~39h) is a
31
+ separate, independently-collected corpus with no evidence of overlap with Common
32
+ Voice -- unlike the Ethiopia/Shona/Lingala/Fulani cases elsewhere in this
33
+ collection, there's no matching-hours or matching-speaker-ID signal here. It remains
34
+ a candidate for a future top-up if more Luganda hours are wanted.
35
+
36
+ All audio is standardized to **16 kHz mono FLAC** (lossless), 1–30 second clips.
37
+ Source audio is MP3, decoded directly via `soundfile` (no external decoder needed).
38
+
39
+ ## Format
40
+
41
+ The dataset ships as **WebDataset-style tar shards** (`shards/shard-00000.tar` …, ~1 GB
42
+ each, one `{key}.flac` file per clip) plus a single manifest (`manifest.parquet` /
43
+ `manifest.jsonl`):
44
+
45
+ | Column | Description |
46
+ |---|---|
47
+ | `key`, `shard` | which tar file + entry holds this clip's audio |
48
+ | `text` | transcript (native Luganda script), from Common Voice's `sentence` field |
49
+ | `duration` | seconds |
50
+ | `source` | always `common_voice` |
51
+ | `dataset_id` | always `0` |
52
+ | `split` | `train` / `val` (250 clips held out for evaluation) |
53
+ | `speaker_id` | Common Voice's `client_id` |
54
+ | `gender` | speaker-reported gender where available (Common Voice's `gender` field, e.g. `male_masculine`/`female_feminine`) |
55
+ | `dbfs`, `clip_ratio`, `sil_ratio` | cheap DSP quality proxies: loudness, fraction of clipped samples, fraction of near-silent frames |
56
+ | `has_disfluency` | always `false` |
57
+
58
+ ## Usage
59
+
60
+ ```python
61
+ from huggingface_hub import hf_hub_download
62
+ import pandas as pd, tarfile, io, soundfile as sf
63
+
64
+ mp = hf_hub_download("Professor/luganda-speech-data", "manifest.parquet", repo_type="dataset")
65
+ df = pd.read_parquet(mp)
66
+
67
+ row = df.iloc[0]
68
+ shard_path = hf_hub_download("Professor/luganda-speech-data", f"shards/{row.shard}", repo_type="dataset")
69
+ with tarfile.open(shard_path) as tar:
70
+ audio_bytes = tar.extractfile(f"{row.key}.flac").read()
71
+ arr, sr = sf.read(io.BytesIO(audio_bytes))
72
+ ```
73
+
74
+ The tar shards are also directly readable by the [`webdataset`](https://github.com/webdataset/webdataset)
75
+ library for streaming training pipelines.
76
+
77
+ ## Intended use & limitations
78
+
79
+ Built for **Luganda TTS/ASR research**, in particular as finetuning data for a
80
+ multilingual TTS model that doesn't natively support Luganda. Common Voice audio is
81
+ crowdsourced (variable recording equipment/environment across contributors) but
82
+ validated by community review, so quality is generally clean read speech. This is a
83
+ **research aggregation**; usage should respect Common Voice's own terms.
84
+
85
+ ## License
86
+
87
+ CC0 1.0 Universal (public domain dedication), per the upstream Common Voice release.
88
+
89
+ ## Acknowledgments
90
+
91
+ Deep thanks to **Mozilla Common Voice** and its Luganda-speaking contributors for
92
+ the source corpus, and to the **Mozilla Data Collective** for hosting it.
93
+
94
+ This dataset was pooled by **Victor Olufemi and LyngualLabs** as part of the
95
+ [AfroNet](https://github.com/osinkolu/afronet-tts-data) multi-language TTS data effort.