# Lero Research

> How Lero's Brazilian voice works inside: acoustics, prosody, PyTorch and infrastructure, with something to hear or play with in every post.

- [A new voice without touching every weight](https://lero.studio/research/en/baixo-rank.md) (November 9, 2026, PyTorch): Low-rank adapters: train under 1% of the numbers and keep the same speed.
- [Attention and the next token](https://lero.studio/research/en/atencao.md) (November 6, 2026, PyTorch): A language model writes one token at a time, looking only backwards.
- [More voices per GPU](https://lero.studio/research/en/gpu.md) (November 4, 2026, Infrastructure): CUDA graph, MPS and batching: how we doubled real-time voices on the same GPU.
- [From audio to tensor](https://lero.studio/research/en/tensores.md) (November 2, 2026, PyTorch): Every audio clip that enters a model becomes a tensor. Its shape decides memory and speed.
- [From sound to tokens](https://lero.studio/research/en/codec.md) (October 31, 2026, Acoustics): One second of speech has tens of thousands of numbers. A voice model does not write them one by one.
- [Same sentence, only the melody changes](https://lero.studio/research/en/melodia.md) (October 28, 2026, Prosody): In a question, the voice rises at the end. And the end of the sentence is where generated voices fail most.
- [Same note, three sounds](https://lero.studio/research/en/timbre.md) (October 22, 2026, Acoustics): Three sounds with the same 220 Hz note. What makes each one sound different is timbre.
- [Inside the sound](https://lero.studio/research/en/espectrograma.md) (October 17, 2026, Acoustics): The same speech seen as an image: time across, frequency up, energy as color.
- [Sound as numbers](https://lero.studio/research/en/onda.md) (October 15, 2026, Acoustics): Before any model, a voice is a list of numbers. Take a close look.