pingo-ptbr / NOTICE
itxtoledo's picture
Add pingo-ptbr: 0.87 MB pt-BR TTS (MLX + ONNX + web demo)
27d8618 verified
Raw History Blame Contribute Delete
1.12 kB
Copyright 2026 SemFence (Gustavo Toledo)
Criado pela SemFence (Gustavo Toledo) - https://semfence.com
Created by SemFence (Gustavo Toledo) - https://semfence.com
pingo-ptbr is a compact edition of the pivo-ptbr family (https://huggingface.co/semfence/pivo-ptbr): a Brazilian Portuguese
text-to-speech model whose weights are stored quantized (5 to 8 bits per weight) in a single file.
The acoustic model (pingo-ptbr.pt and the acoustic part of pingo-ptbr.bin) was trained from scratch by SemFence. Its design
follows PiVo-1.4M, Copyright 2026 Empero Research (Leon Lehmann), Apache License 2.0 (https://huggingface.co/empero-ai/PiVo-1.4M),
with a narrower width. No PiVo acoustic weights are used.
The vocoder part of pingo-ptbr.bin (and of pingo-ptbr-w96.bin) is the PiVo vocoder, converted from the FARGAN vocoder of
Xiph/Opus (BSD-3-Clause), with its weights stored at 5 to 8 bits per weight instead of 32. Its notices are in LICENSES/.
The acoustic weights are licensed under CC BY 4.0 (credit required: SemFence, Gustavo Toledo).
The code is licensed under Apache-2.0. The vocoder keeps its BSD-3-Clause license.