GhanaNLP/ghana-speech-english-ipa-latents-data-prep

Data prep for the VoxCPM-2 Ghanaian TTS finetune: phoneme ASR, AudioVAE latent precompute, and the IPA latent datasets it publishes # ghana-speech-english-ipa-latents-data-prep Data prep for the **VoxCPM-2 Ghanaian TTS** finetune: turn raw speech corpora into **precomputed AudioVAE latents paired with IPA transcripts**, so training never decodes audio or runs the VAE. Produces four datasets on the Hub — the same 418,265 clips in two forms: | dataset | contents | subsets | clips | hours | |---|---|---|---|---| | `ghanaopendata/voxcpm2-ghana-ipa-latents` | AudioVAE latents | 42 Ghanaian languages | 365,295 | ~748 | | `ghanaopendata/voxcpm2-english-ipa-latents` | AudioVAE latents | Ghanaian English | 52,970 | ~201 | | `ghanaopendata/ghana-speech-ipa` | **audio** + text + IPA | 42 Ghanaian languages | 365,295 | ~748 | | `ghanaopendata/ghana-english-speech-ipa` | **audio** + text + IPA | Ghanaian English | 52,970 | ~201 | The **latent** datasets are consumed by `train_voxcpm_finetune.py` with `train_manifest: ghana-latents` — no audio decoding or VAE pass at train time. The **audio** datasets are the same clips with real FLAC audio, for listening, checking transcripts, or training anything that is not VoxCPM-2. Two deliberate differences between the pairs: - the latents keep `split` as a **column**, because `train_voxcpm_finetune.py` filters on `r["split"]`; the audio datasets use real `train`/`validation` **splits** - in the latents `text` holds the **IPA** (the trainer re-tokenizes `text` directly, so graphemes there would silently train on the wrong units) with the orthography in `raw_text`; the audio datasets name them `ipa` and `text` ## Pipeline ``` ghana-speech (audio parquets) │ ├─ ghana-ipa-asr batch ───────────► ghana-speech-ASR-IPA2/ /*.parquet │ GPU CTC phoneme ASR at ~2100x realtime, IPA read off the actual audio. │ Driver: asr2.sh. (transcribe_ghana_gpu_asr.py is the superseded version — │ it skipped input normalisation; see below.) │ ├─ merge_asr2_ipa.py ─────────────► ghana-speech-IPA/ /*.parquet │ (shard, row) → id → IPA lookup, replaces the old g2p labels │ ├─ pr …