v0.8.4

speech, speech.cpp's one executable (tts, asr, voice, info, devices, models, pull, rm, quantize, serve and worker), and the shared library libspeech with its C API, speech.h, in one archive for each of macOS arm64 (Metal), Windows x64 (Vulkan) and Linux x64 (Vulkan, and the CPU alone; glibc 2.34 or later). The README lists the model families and where their converted weights are on Hugging Face.

What's Changed

• Document that Irodori-TTS MF can garble long sentences and how to avoid it by @nyosegawa in #101
• Say that speech serve's --add-voice takes a WAVE file as well as a voice file by @nyosegawa in #102
• Load a model's tensors from its file 16 MB at a time by @nyosegawa in #105
• Check Irodori-TTS's refusal of a speed of 4 in the voice none, and count each request's own samples by @nyosegawa in #104
• Let speech serve and speech worker take request options as their models' defaults, per task with a prefix by @nyosegawa in #103
• Say that a program frees its models before it ends, as Metal aborts one that does not by @nyosegawa in #106
• Compare the live panel's text with the API's for the audio the page sent by @nyosegawa in #107
• Skip the build for a change that touches only Markdown, docs/ or skills/ by @nyosegawa in #108
• Add README.ja.md, the README in Japanese, kept in step with README.md by @nyosegawa in #109
• Build on ggml from a fork that carries a fix and a test for each defect of its upstream tag by @nyosegawa in #111
• Offer every weight type of each model, and fetch Q8_0 for reazonspeech-v2 and parakeet-ja by @nyosegawa in #110
• Raise the version to 0.8.4 by @nyosegawa in #112

Full Changelog: v0.8.3...v0.8.4