bookbot/OpenBible_Swahili

Dataset Card for OpenBible Swahili

OpenBible Swahili is a verse-level segmented audio of Swahili: Biblica® Open Kiswahili Contemporary Version (Neno). This dataset/project is inspired by masakhane-io/bibleTTS and coqui-ai/open-bible-scripts, with a twist of applying the more recent Massively Multilingual Speech (MMS) for audio alignment.

Dataset Details

Dataset Description

For a more detailed description, please refer to OpenBible-TTS.

We followed the CTC Forced Alignment API Tutorial graciously provided by PyTorch. We used the Massively Multilingual Speech (MMS) model to align the audio to the text.

Like the challenges explained in the MMS paper, we also faced the same noisy audio alignment issues:

• Chapter introduction narration
• Occasional verse number reading
• Digits/number handling

While masakhane-io/bibleTTS proposes a solution by manually (1) inserting chapter introduction transcript and (2) spells out numbers, we decided to use a mix of MMS' method by (1) inserting * token for additional speech (e.g. chapter introduction, verse number reading) and (2) converting digits to words using num2words, if available.

The aligned verse text and audio are then segmented into verse-level segments for TTS model training.

As proposed in §3.1.5 of MMS, we also implemented a length-normalized probability difference filtering to remove noisy alignments based on the following equation:

$$\frac{1}{T}[\log P(Y^{\text{aligned}}\mid X)-\log P(Y^{\text{greedy}}\mid X)]\backslash frac\{1\}\{T\}\; \backslash left[\backslash log\; P\backslash left(Y^\{\backslash text\; \{aligned\}\}\; \backslash mid\; X\backslash right)-\backslash log\; P\backslash left(Y^\{\backslash text\; \{greedy\}\}\; \backslash mid\; X\backslash right)\backslash right]$$

where $TT$ is the length of the audio, $P(Y^{\text{aligned}}\mid X)P\backslash left(Y^\{\backslash text\{aligned\}\}\; \backslash mid\; X\backslash right)$ is the probability of the forced-alignment path, and $P(Y^{\text{greedy}}\mid X)P\backslash left(Y^\{\backslash text\{greedy\}\}\; \backslash mid\; X\backslash right)$ is the probability of the greedy sequence.

Like MMS, we select −0.2 as the default threshold and choose samples with scores greater than this threshold. We call the resultant subsets clean, which you can load as shown in the Usage section.

⚠️ 2 Chronicles 13-14 (2CH_013, 2CH_014) have corrupt source audio. In the non-clean subsets, these chapters might be problematic for training!

We recommend reading the MMS paper for better understanding.

• Homepage: https://github.com/bookbot-hive/OpenBible-TTS/
• Repository: https://github.com/bookbot-hive/OpenBible-TTS/
• Point of Contact: Wilson Wongso

Usage

from datasets import load_dataset

# all 66 books
dataset = load_dataset("bookbot/OpenBible_Swahili", trust_remote_code=True)
# all 66 books, clean subset
dataset = load_dataset("bookbot/OpenBible_Swahili", "clean", trust_remote_code=True)
# only Genesis
dataset = load_dataset("bookbot/OpenBible_Swahili", "GEN", trust_remote_code=True)
# only Genesis, clean subset
dataset = load_dataset("bookbot/OpenBible_Swahili", "GEN_clean", trust_remote_code=True)

Dataset Structure

Data Instances

{
    "id": "GEN_001_001",
    "verse_id": "GEN 1:1",
    "audio": {
        "path": "GEN/GEN_001/GEN_001_001.wav",
        "array": array([0.0010376, 0.0007019, 0.00015259, ..., 0.00817871, 0.00482178, 0.01132202]),
        "sampling_rate": 44100,
    },
    "verse_text": "Hapo mwanzo Mungu aliumba mbingu na dunia.",
    "transcript": "hapo mwanzo mungu aliumba mbingu na dunia",
}

Data Fields

• id: a unique identifier for the verse, formatted as BOOK_CHAPTER_VERSE.
• verse_id: the verse identifier, formatted as BOOK CHAPTER:VERSE.
• audio: a dictionary containing the path to the audio file, the decoded audio array, and the sampling rate.
  • path: the path to the audio file.
  • array: the decoded audio array.
  • sampling_rate: the sampling rate of the audio.
• verse_text: the verse text, with proper capitalization and punctuations.
• transcript: the lowercased, punctuation-removed version of verse_text, used during forced-alignment.

Data Splits

USX Code	Book Name	#verses	#verses (clean)
GEN	Genesis	1533	1498
EXO	Exodus	1213	1070
LEV	Leviticus	859	813
NUM	Numbers	1288	1188
DEU	Deuteronomy	959	940
JOS	Joshua	658	585
JDG	Judges	618	610
RUT	Ruth	85	83
1SA	1 Samuel	810	799
2SA	2 Samuel	695	672
1KI	1 Kings	816	773
2KI	2 Kings	719	597
1CH	1 Chronicles	942	859
2CH	2 Chronicles	822	692
EZR	Ezra	280	226
NEH	Nehemiah	406	353
EST	Esther	167	166
JOB	Job	1070	1050
PSA	Psalms	2461	2367
PRO	Proverbs	915	893
ECC	Ecclesiastes	222	221
SNG	Song of Songs	117	114
ISA	Isaiah	1292	1223
JER	Jeremiah	1364	1297
LAM	Lamentations	154	149
EZK	Ezekiel	1273	1258
DAN	Daniel	357	335
HOS	Hosea	197	194
JOL	Joel	73	70
AMO	Amos	146	144
OBA	Obadiah	21	20
JON	Jonah	48	48
MIC	Micah	105	103
NAM	Nahum	47	46
HAB	Habakkuk	56	55
ZEP	Zephaniah	53	53
HAG	Haggai	38	37
ZEC	Zechariah	211	211
MAL	Malachi	55	55
MAT	Matthew	1071	1004
MRK	Mark	678	557
LUK	Luke	1151	1027
JHN	John	879	856
ACT	Acts	1007	916
ROM	Romans	433	416
1CO	1 Corinthians	437	404
2CO	2 Corinthians	257	252
GAL	Galatians	149	145
EPH	Ephesians	155	152
PHP	Philippians	104	99
COL	Colossians	95	89
1TH	1 Thessalonians	89	86
2TH	2 Thessalonians	47	47
1TI	1 Timothy	113	109
2TI	2 Timothy	83	83
TIT	Titus	46	42
PHM	Philemon	25	24
HEB	Hebrews	303	278
JAS	James	108	108
1PE	1 Peter	105	91
2PE	2 Peter	61	56
1JN	1 John	105	103
2JN	2 John	13	13
3JN	3 John	15	15
JUD	Jude	25	24
REV	Revelation	404	370
	Total	31103	29233

Additional Information

Licensing Information

CC-BY-SA 4.0