github gradium-ai/xn-ptts v0.6.0

3 hours ago

Phonon 0.6.0

Changes since 0.5.0.

  • macOS 27: Metal works again. The new Metal compiler rejected one of the matrix-multiply kernels, so --device metal, and the default --device auto with it, failed before speaking (#302). Fixed in xn 0.2.13 (gradium-ai/xn#132, #309).
  • GPU fallback: with --device auto, a GPU that fails to start now gives way to the CPU with a one-line warning, instead of failing the whole load. Python raises it as a UserWarning. A device chosen by name still fails with the full error (#310).
  • Thirteen more languages: --lang takes ar, hi, ko, pl, sv, fi, no, it, nl, ja, zh, da and tr, in every frontend. They get the character pass with their own spoken forms for symbols, and Japanese is spaced one character at a time as its checkpoints were trained; the word rules (numbers, currency, emails, ...) stay with en, fr, de, es and pt. These codes need a checkpoint trained on the language (#304, #305, #306).
  • Core ML: the exported Mimi decoder attends to the same window as ptts (#281) and uses the checkpoint's GELU (#282), its size comes from the checkpoint's config (#283), and the smaller Pocket TTS models export too (#284).
  • Fixes:
    • Hugging Face Hub: HF_TOKEN, HF_HOME, HF_HUB_CACHE, HF_ENDPOINT and HF_HUB_OFFLINE are read the way huggingface_hub reads them, and on Windows the cache and the hf auth login token are found under the user's profile (#308).
    • CLI: log lines go to stderr, so -o /dev/stdout carries only the audio (#293).
    • Metal and CUDA: one wait on the device per frame instead of two (#280).
  • QNN: the runner builds on Windows ARM64 (#273), and the 31m bundle has a Snapdragon X Elite context (#300).
  • Rust API (breaking): ptts::preprocess::Lang has thirteen new variants and is now #[non_exhaustive], so a match on it outside ptts needs a _ arm; Lang::ALL is a slice (#304, #312). This is why the version is 0.6.0. From now on, new languages are not breaking.

The attached phonon-31m-en-coreml.zip is exported anew with this release's Core ML fixes. phonon-31m-en-qnn.zip is the one attached to 0.5.0, with the Snapdragon X Elite context.

Don't miss a new xn-ptts release

NewReleases is sending notifications on new releases.