Indonesian Accented Read Speech Corpus
Audio-only Indonesian (id) speech-recognition dataset built by merging 11 audio-visual lip-reading corpora from the Mozilla Data Collective. The video streams of the originals have been discarded; only the audio track of each recording is included, paired with its transcript.
At a glance
| |
|---|
| Utterances | 37,351 |
| Total audio | 60.52 h |
| Source datasets | 11 |
| Audio | OPUS, mono, 16000 Hz, 48 kbit/s |
| Splits | train 25,460 (41.82 h) · dev 6,134 (9.95 h) · test 5,757 (8.75 h) |
| Dialects | Lumajang (5,768), Multidialect (3,934), Tulungagung (3,459), Jember (3,457), Gresik (3,434), Semarang (3,250), Banyumasan (2,898), Solo (2,884), Betawi (2,883), Bogor (2,859), Yogya (2,525) |
| Split strategy | speaker-disjoint, deterministic hash |
| Cross-source duplicates removed | 0 |
| Items skipped (no audio / filtered / failed) | 0 |
| Built | 2026-11-04 |
Licence
The bundle is distributed under the MDC Data Licence Agreement 1.0 — see LICENCE.md and https://community.mozilladatacollective.com/mdc-data-licence-agreement-1-0/. It is made available, including for commercial use, under the provider agreements between Mozilla Data Collective and the owners of each source dataset.
For provenance only, the source datasets' individual public MDC listings carry the licences below. Those licences apply to the individual listings, not to this bundle. Original notices are preserved in sources// and listed in ATTRIBUTION.md:
- CC-BY-NC-SA-4.0: 6 dataset(s)
- CC-BY-SA-4.0: 5 dataset(s)
Language and speakers
All transcripts are in Indonesian. Each speaker folder in the sources is labelled Jawa or Indonesia plus a regional dialect (e.g. Jawa_Solo Dialect_53). The transcripts give no sign that the Jawa speakers read in Javanese: they read the same Indonesian prompts as everyone else. The label therefore describes the speaker’s background and accent (the column speaker_group), not the language of the recording.
Speaker groups across the bundle (by audio duration): Jawa 78.5% (47.52 h), Indonesia 21.5% (12.99 h)
| Dataset | Jawa (audio) | Indonesia (audio) | Utterances with Javanese words |
|---|
| Bahasa Indonesia Audio Visual Speech Dataset | 85.0% | 15.0% | 0 |
| INALVS (Indonesian Lip-reading & Visual Speech Dataset) | 86.5% | 13.5% | 0 |
| Indonesia Lip-Record Speech Dataset | 85.7% | 14.3% | 0 |
| Indonesian Lip-Record Multimodal Corpus | 69.9% | 30.1% | 7 |
| Speech Corpus of Indonesian Lip-Record | 85.4% | 14.6% | 1 |
| Indo-Labial Multimodal Corpus | 87.4% | 12.6% | 0 |
| Bahasa Indonesia Lip-Reading Dataset | 68.2% |
Speakers. The sources do not give speaker IDs. Their numbered folders (1,474 in total) are recording sessions, not speakers: each source's datasheet describes about six speakers. Each L1-group/dialect label corresponds to one speaker and keeps the same session naming in every source, so speakers are inferred from that label (the Lumajang and Multidialect labels each cover two speakers, told apart by their session-folder style; a voice-embedding check confirmed these splits). This gives 13 inferred speakers across the bundle; the same person therefore appears in several sources. The original folder is kept in the session column.
| # | Speaker | Gender | Age | Utterances | Audio | Sources |
|---|
| 1 | Jawa-Tulungagung | male | 20-29 | 3,459 | 6.48 h | 6 |
| 2 | Jawa-Jember | male | 20-29 | 3,457 | 5.75 h | 6 |
| 3 | Jawa-Yogya | male | 20-29 | 2,525 | 5.45 h | 6 |
| 4 | Jawa-Lumajang-2 |
Demographics. Gender: male 9 speakers (42.10 h, 69.6% of audio), female 4 speakers (18.41 h, 30.4% of audio). Age: 20-29. Recorded per speaker in the gender and age columns; supplied separately for the inferred speakers, not taken from the source files.
The bundle has 3,663 distinct sentences across 37,351 utterances (each read 10 times on average, by different speakers and sources). Splits are speaker-disjoint, so no speaker appears in more than one split.
Because every speaker read from the same pool of prompts, a speaker-disjoint split cannot also be sentence-disjoint: dev and test contain unseen voices reading sentences that also occur in train. Test results therefore measure generalisation to new speakers, not to new text.
Source datasets
Layout
audio//<utt_id>.opus
train.tsv dev.tsv test.tsv all.tsv # Common Voice-style, tab-separated
train.jsonl dev.jsonl test.jsonl # NeMo-style manifests
sources//datasheet.json # datasheet as returned by the MDC API
sources//original_docs/ # README/licence files found in the archive
contributions.tsv # per-source utterances, seconds, share of audio
duplicates.tsv skipped.tsv bundle_manifest.json
TSV columns: path, sentence, duration, split, speaker, session, dialect, speaker_group, age, gender, javanese_words, source_dataset, source_licence, original_file, utt_id. path is relative to the bundle root.
Loading
import pandas as pd, soundfile as sf
train = pd.read_csv('train.tsv', sep='\t', quoting=3)
audio, sr = sf.read(train.path[0])
from datasets import load_dataset, Audio
ds = load_dataset('csv', data_files={'train': 'train.tsv', 'validation': 'dev.tsv', 'test': 'test.tsv'},
delimiter='\t', quoting=3).cast_column('path', Audio(sampling_rate=16000))
Processing
- Each source archive was downloaded from MDC with the
datacollective SDK.
- Transcript TSVs were matched to their media files; rows without a resolvable file or with an empty transcript were dropped.
- Where a source ships a separate audio file per recording it was used directly; otherwise the audio track was taken from the video. Audio was down-mixed to mono and resampled to 16000 Hz. No video frames are kept.
- Transcripts were whitespace-normalised only (case and punctuation are as in the source).
- Clips shorter than 0.3s or longer than 30.0s were removed.
- Byte-identical audio (and byte-identical source files) appearing in more than one source were kept once; see
duplicates.tsv.
- Splits were assigned by hashing the speaker so the same speaker never appears in two splits.
Known limitations
- The speaker pool is small (see “Language and speakers”), and speaker IDs are inferred. With the default sentence-disjoint split, the same speakers appear in train, dev and test, so test results measure performance on known voices. Use
--split-by speaker to hold out whole speakers instead.
- Recordings were made for lip reading, so microphone quality, background noise and level vary between sources.
- Regional/Javanese-influenced pronunciation is present by design.
dialect and speaker_group come from the source TSVs where present, otherwise from the per-speaker folder names (e.g. 001_Indonesia_Semarang Dialect, Jawa_Solo Dialect_53); they are as labelled by the providers and not verified.
- Some session folders carry a letter suffix (e.g.
48a alongside 48); these are treated as further sessions, not separate speakers.
- Contributor names found in the source folder names have been replaced with invented names of the same apparent gender in
session, original_file and the build notes. Transcripts are unchanged.
- Do not attempt to identify the speakers (see the licence).