Datasets:

edwixx
/

bemtevi-tts-data

Tasks:

Modalities:

Formats:

Languages:

Size:

Tags:

brazilian-portuguese

voxcpm2-latents

Libraries:

License:

Dataset card Data Studio Files Files and versions

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Bem-te-vi TTS data

Training data for Bem-te-vi TTS, a small, fast Brazilian Portuguese text-to-speech model.

Source audio: Multilingual LibriSpeech Portuguese (CC BY 4.0), 37533 clips, 160.96 h, 16 kHz.

Path	What
`transcripts.jsonl`	per clip: speaker, book, original MLS text (`orig`), Whisper large-v3-turbo text (`text`, via Kestrel; lowercase, no punctuation), `text_pause` (punctuation from the reader's pauses), word timestamps (`words`), CER of Whisper vs original (`cer`), latent frame count
`latents/shard_XX.pt`	`dict[id -> float16 tensor [T, 64]]`, VoxCPM2 AudioVAE latents at 25 Hz (16 kHz in, decodes to 48 kHz)
`decoder_test/`	original vs VoxCPM2 VAE decode vs EMA Lightning decoder, plus `report.json`
`summary.json`	hours per speaker, CER stats

Downloads last month: 46

Total file size: