Build an Indonesian text-to-speech training dataset in the browser. Audionesia Dataset Generator ships a phoneme-balanced reading corpus, records raw PCM from the microphone, keeps everything in IndexedDB, and exports the dataset/ layout that StyleTTS2 and PocketTTS training expect. No server, no account: open the page and record.

{"hash":"61455d3da0556e62","path":"dataset/audio/budi/clip_0001.wav","text":"Tak seorang pun boleh ditangkap.","phonemes":"taʔ səoraŋ pun boleh ditaŋkap.","duration":3.727,"speaker":"budi"}

What it does

The home page lists datasets as folders. Each dataset has a name, one speaker, a target in hours, and a progress bar of approved recording time. Open a dataset to work in its tabs:

Tab	Job
Rekam	One script at a time, microphone picker, level meter, raw capture with browser audio processing off, silence trim, waveform of the take. Scripts built from your own text come first.
Dengarkan	Review card with play/pause, waveform, Ya / Tidak, re-queue, ASR check; below it every clip of the dataset with status filters, text search, and pages.
Teks sendiri	Paste or drop your own text (.txt, .tsv, .jsonl, sample files to download); it is filtered, phonemized, added to the pool, and its scripts go to the front of the Rekam queue.
Dataset	Progress toward the dataset's target, coverage of phones, diphones, and spelling phenomena, export with a preview of the folder tree, ASR check of all pending clips.
Pengaturan	Trainer presets, duration window, silence and clipping gates, export sample rate and peak level, speech-rate calibration, ASR model.

Scripts are built by a greedy set cover over the pool's coverage units, so the rarest sounds are read first; each script stays within one text source so it reads as one voice. Phonemes come from indo-g2p, and every phoneme string is stored with the library version that produced it.

Before you record

• StyleTTS2 ships an English-only PL-BERT ("it probably does not work very well on other languages", per its README). Train or find an Indonesian PL-BERT before fine-tuning on this dataset. StyleTTS2 also crops training audio to 5 s by default and rejects phoneme strings over 512 tokens, so the StyleTTS2 (5-15 s) preset is the default.
• PocketTTS has no Indonesian model; its new-language path needs an Indonesian tokenizer and a forced aligner (see its training/README.md), and its authors size such a run at 100 hours or more. Pick the PocketTTS (10-30 s) preset in Pengaturan and rebuild scripts before recording for it.
• Licensing. Only Common Voice (CC0) and generated text are free of obligations. Tatoeba text is CC BY 2.0 FR and Wikipedia text is CC BY-SA 4.0; a dataset or model built on them carries attribution and, for BY-SA, share-alike questions. Pengaturan has a CC0-only export mode, and every export writes ATTRIBUTION.md.
• Back up. Browsers can evict local storage. The home page backs up every dataset with its clips' master audio to a folder or ZIP and restores them on any machine.

Quickstart

Prerequisites: Bun 1.4 or newer. Google Chrome only for bun run smoke and bun run banner.

git clone git@github.com:snowfluke/audionesia-dataset-gen.git
cd audionesia-dataset-gen
bun install              # also installs the git hooks
bun run dev              # http://localhost:5173

The first load seeds the bundled pool (30,071 phonemized entries) into IndexedDB and builds the scripts; a redeployed pool reseeds itself. Create a dataset (name, speaker name, target hours, optional gender, age range, dialect, microphone, and the consent checkbox), allow the microphone, press Space.

Verify a change:

bun run complete-check   # type-check, lint, fmt, test, build
bun run smoke            # Playwright drives Chrome through record, review, export, backup; needs bun run dev

Export layout

Each export covers one dataset, so one speaker. Two datasets exported into the same folder overwrite each other's speakers.jsonl and manifest.json; use one folder per dataset or merge the JSONL files yourself.

dataset/
  audio/<speaker>/clip_0001.wav    # mono 16-bit, 24 kHz by default, DC removed, peak -3 dBFS
  speakers.jsonl                   # {"hash","path","text","phonemes","duration","speaker"}
  metadata.jsonl                   # Hugging Face audiofolder (file_name, text, phonemes, speaker, duration, split)
  styletts2/train_list.txt         # path|phonemes|speaker_id, plus val_list.txt and sentence-level OOD_texts.txt
  pocket-tts/train.jsonl           # {"path","duration","transcript"}, plus valid.jsonl
  manifest.json                    # sample rate, counts, g2p versions, speakers with metadata, warnings count
  export-warnings.txt              # every clip a trainer would reject, with the reason
  ATTRIBUTION.md                   # license and credit for every text source used

hash is the first 16 hex characters of the SHA-256 of the exported WAV. Train/validation membership is a hash of the clip id, so it never changes between exports. Only approved clips export; clips shorter than the minimum, or whose text is not CC0 in CC0 mode, are skipped and listed. On Chromium the app writes straight into a folder you pick and prunes WAVs of clips that no longer exist; other browsers get a ZIP.

Quality gates

• Clipped takes cannot be saved (switchable). Each take shows its peak and the signal-to-noise ratio of the leading silence.
• Takes stop themselves after 90 s; a disconnected microphone or a paused audio context aborts the take with a message.
• The in-browser ASR check runs Whisper (onnx-community/whisper-base by default, WebGPU when available) on a clip or on every pending clip and flags transcripts that differ from the script by more than 20%.

Corpus

The pool under public/corpus/ is generated, committed, and rebuilt with:

bun run corpus:common-voice    # 6,344 CC0 sentences from the Common Voice repository
bun run corpus:tatoeba         # 28,192 CC-BY sentences with per-sentence attribution
bun run corpus:wikipedia 2000  # random id.wikipedia intros, CC-BY-SA
bun run corpus:news 200        # feeds listed in corpus/sources.json, read through the ax CLI
bun run corpus:llm 5           # Claude-written paragraphs for the least-covered units (ANTHROPIC_API_KEY)
bun run corpus:build           # normalize, filter, dedup, phonemize -> pool.jsonl, index.json, ATTRIBUTION.md

The build drops fragments under four words and lines that indo-g2p reads mostly as English. corpus/raw/ is not committed. The current pool holds 30,071 entries, 989 coverage units, and about 52 hours of reading text; public/corpus/ATTRIBUTION.md lists what each source requires of a derived dataset.

Repository map

audionesia-dataset-gen/
├── src/
│   ├── app.tsx  main.tsx        # shell: header, home page or workspace tabs, help dialogs
│   ├── components/              # Kumo-styled Solid components, one per file
│   ├── features/                # workspaces, record, review, write, dataset, settings, library
│   ├── lib/                     # audio, asr, backup, corpus, db, export, g2p, text: pure logic and repositories
│   └── workers/                 # G2P worker, script-builder worker, ASR worker, recorder worklet
├── public/corpus/               # generated pool (committed)
├── scripts/corpus/              # fetch and build scripts
├── scripts/smoke.ts             # browser smoke test
├── tests/                       # bun test, mirrors src/lib
├── docs/                        # coding standard, review checklist
└── .claude/skills/              # solidjs-2 and kumo-design skills for agents

Stack: SolidJS 2.0 (release candidate, pinned), Vite 8, Tailwind CSS 4 with Cloudflare Kumo tokens, idb, zod, fflate, indo-g2p, @huggingface/transformers (ASR only); Bun, oxlint, oxfmt, lefthook.

Documentation

• CLAUDE.md: the operating manual, invariants, and every command.
• docs/CODING_STANDARD.md and docs/CODE_REVIEW_CHECKLIST.md.
• .claude/skills/solidjs-2/: Solid 2.0 API and migration reference.

Deploy

.github/workflows/pages.yml builds and publishes dist/ on every push to main. Enable it once under Settings > Pages > Source: GitHub Actions.

License

MIT. Text sources keep their own licenses: Common Voice sentences are CC0, Tatoeba sentences are CC BY 2.0 FR, Wikipedia text is CC BY-SA 4.0, VOA Indonesia reporting is public domain. The export writes ATTRIBUTION.md next to the audio so a published dataset carries them.