Synthetic Javanese–Indonesian Conversational Speech

Synthetic Javanese–Indonesian conversational speech (Lost in Transcription, Indonesian–Javanese track)

13,469 fully synthetic utterances (61.2 h) of informal Javanese, Indonesian and code-mixed Javanese–Indonesian speech. The dataset was created by participant inyongkhafid to train ASR models for the Lost in Transcription challenge (Mozilla Data Collective / DrivenData, 2026), and is published on MDC because the competition requires external training data to be published on MDC. No real person was recorded for this dataset. All texts are machine-generated, and all audio is produced by a text-to-speech model.

Contents

Subsets (generation batches): sintetis2 1,660 clips (8.3 h) · sintetis3 7,445 clips (35.5 h) · sintetis5 4,364 clips (17.4 h)

How it was made

1. Text. Short conversational monologues were generated locally with Gemma 4 26B-A4B (Q4_K_M GGUF, llama.cpp), prompted with a speaker persona, a topic, and two style examples taken from the competition development transcripts. A rule-based filter removed malformed texts. One batch (sintetis5) is Indonesian conversational text, partly rewritten deterministically. Overlap check against the development transcripts: 0 of 2,611 unique texts share an 8-word sequence; 1 text (4 clips) shares one generic 6-word phrase about a price.
2. Voice. VoxCPM2 (openbmb/VoxCPM2, Apache-2.0) with a LoRA adapter fine-tuned by the uploader on 4,000 utterances of OpenSLR 41 (CC BY-SA 4.0) and 2,000 utterances of the MDC dataset TTS Central Javanese (CC BY-SA 4.0). Voices were chosen with textual persona descriptions (e.g. "young man chatting casually, thick Javanese accent"); no reference recording was used to clone a specific voice.
3. Augmentation. Each clip was rendered clean, through an Opus codec, or with a tempo change, then resampled to 16 kHz mono.

How it was used

Together with real corpora (Jember Javanese Spontaneous Speech Corpus, OpenSLR 35/41, Common Voice, MDC podcasts, TTS Central Javanese), these clips fine-tuned the members of the Indonesian–Javanese competition ensemble: Qwen3-Omni-30B-A3B, Qwen3-ASR-1.7B, omniASR-LLM-7B and Whisper large-v3 LoRA. The 4,519 original TTS Central Javanese recordings that were also used in training are not included here; they are available on MDC.

Licence and attribution

CC BY-SA 4.0. The voices derive from a model adapted on OpenSLR 41 and TTS Central Javanese, both CC BY-SA 4.0. The owner of TTS Central Javanese granted permission for this use and for publishing the synthetic data (30 September 2026). No additional terms are added.

• OpenSLR 41: Sodimana, K., Pipatsrisawat, K., Ha, L., Jansche, M., Kjartansson, O., De Silva, P., & Sarin, S. (2018). A Step-by-Step Process for Building TTS Voices Using Open Source Data and Framework for Bangla, Javanese, Khmer, Nepali, Sinhala, and Sundanese. Proc. SLTU 2018, 66–70.
• TTS Central Javanese — Fithrotunnisa, R. (2026). Mozilla Data Collective, dataset cml5bn4k900aame07u0rwidcg.

Ethics

The speech is synthetic, but the TTS adapter was trained on recordings of real speakers, so some voices may resemble them. Please do not use this dataset for speaker identification, voice cloning or impersonation (a request, not a licence condition). The transcripts contain no personal data.

Samples

• s2_0013_p2 — Pas lagi di pasar... terus aku nyoba beli cenil yang warnanya ijo gitu. Terus pas lagi mangan... e... teksturnya kenyal banget sampe gigiku kayak nyangkut. Aduh, susah banget... eh, maksudnya malah jadi susah nelen. Gak estetik banget deh.
• s2_0076_p5 — Mbiyen pas musim tanam, rame banget lho, gitu... kabeh-kabeh bareng-bareng ngolah sawah. Saiki kan... ya gitu, wong-wong wis mulai tinggal neng kota, sawah yo diurus karo sing wis tuwa-tuwa wae. Dadi koyo... sedih gitu.
• s3_r5f1228_1364_p1 — Akhire biji ujianne wis metu. Alhamdulillah, angkane apik-apik banget, gak elek. Anakku seneng banget, terus njaluk tuku bakso... e... bakso urat neng pasar. Yo nggo ngerayakke merga wis lulus ujian sing angel kuwi.
• s3_r5j1x0495_451_p4 — Dalan-dalan nang daerah Jawa Tengah kuwi... e... kadhang akèh bolongane, ya. Dadine nek liwat wengi-wengi, ngeri banget ngrasakne guncangane. Terus... ngono, stir-e dadi sering goyang-goyang dewe. Kaya... eh, maksudé setire dadi abot banget nek kena lubang ngono.
• s3_r5j2x0106_090_p0 — Terus si Agus crito, angin sore iki kan semilir banget... e... penak nggo miberke layangan. Dheweke ngomong, nek pengen layangane mumbul dhuwur, awakdewe kudu pinter ngulur benang... terus ojo sampek gelasane adoh banget, ben ora angel narike.
• s5_r5f0816_970_p0 — Aku pengen deh, gaya berpakaianku kayak artis yang lagi naik daun itu. Yang bajunya selalu kelihatan mewah. Nggak perlu yang mahal sih, yang penting, ya, yang mirip-mirip dikit lah. Biar kalau lagi jualan di pasar, penampilanku, jadi lebih segar dilihat orang.