Lero

Back to Lero

Research

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

  1. PyTorch · 2 min read

    A new voice without touching every weight

    Low-rank adapters: train under 1% of the numbers and keep the same speed.Read

  2. PyTorch · 2 min read

    Attention and the next token

    A language model writes one token at a time, looking only backwards.Read

  3. Infrastructure · 2 min read

    More voices per GPU

    CUDA graph, MPS and batching: how we doubled real-time voices on the same GPU.Read

  4. PyTorch · 2 min read

    From audio to tensor

    Every audio clip that enters a model becomes a tensor. Its shape decides memory and speed.Read

  5. Acoustics · 2 min read

    From sound to tokens

    One second of speech has tens of thousands of numbers. A voice model does not write them one by one.Read

  6. Prosody · 2 min read

    Same sentence, only the melody changes

    In a question, the voice rises at the end. And the end of the sentence is where generated voices fail most.Read

  7. Acoustics · 2 min read

    Same note, three sounds

    Three sounds with the same 220 Hz note. What makes each one sound different is timbre.Read

  8. Acoustics · 2 min read

    Inside the sound

    The same speech seen as an image: time across, frequency up, energy as color.Read

  9. Acoustics · 2 min read

    Sound as numbers

    Before any model, a voice is a list of numbers. Take a close look.Read