michsethowusu/VoxCPM-Synthetic-Data-Generator

Multi-user synthetic speech data generator Space for VoxCPM (Twi/Ewe/Dagbani) — task queue, resume, retries, email notifications --- title: VoxCPM Synthetic Data Generator emoji: 🏭 colorFrom: green colorTo: blue sdk: gradio sdk_version: 5.46.0 app_file: app.py pinned: false hf_oauth: true --- # VoxCPM Synthetic Speech Data Generator — Twi · Ewe · Dagbani Generates synthetic TTS training data with ghananlpcommunity/voxcpm-twi-ewe-dagbani-full, voice-cloning the built-in male/female reference speakers. ## How it works 1. Sign in with Hugging Face (usage is logged to `_usage/usage_log.jsonl` in the output repo). 2. Pick a dataset — all `ghananlpcommunity` datasets are pre-listed (searchable), or type any custom id. "Inspect dataset" detects the string columns: auto-selected if there is exactly one, otherwise you choose. 3. **Preview 5 samples** to hear the result before committing to a big run. 4. Choose voices: all male, all female, or a custom male/female split (deterministic per row, so resumed runs assign the same voice). 5. Set target hours of audio (capped at 10 h) and a GPU profile — the profile sets how many parallel VoxCPM instances run off a shared work queue (~4.5 GB VRAM each). Live status shows throughput and ETA. 6. Output goes to the public dataset repo **ghananlpcommunity/voxcpm-twi-ewe-dagbani-synthetic** under ` - /`: - `audio/ _ .wav` — 16 kHz mono, silence-trimmed (raw files, not parquet) - `manifest.jsonl` — id, file, text, gender, speaker, duration per row - `progress.json` — resume state 7. Commits are batched (default 200 files / 10 min) to stay under HF's 128 commits/hour limit. ## Task queue Tasks from any signed-in user queue up and run one at a time while the Space is alive. The live table shows status, progress, ETA, attempts and a **link to each task's data folder** (partial results appear there as commits land). The queue snapshot is persisted to `_usage/tasks_state.json`, so unfinished tasks resume automatically after a Space restart. ## Retries & notifications A failing attempt (GPU exhaustion, stall >15 min, 20 consecutive row errors) is retried up to 3× …