v0.7.1

speech, speech.cpp's one executable (tts, asr, voice, info, devices, 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.

Changes

• The Qwen3 decoder is faster on a GPU, for Qwen3-ASR and Qwen3-TTS alike, with the same texts and the same checks:
  • It attends with flash attention where the device computes it, chosen at load and logged; elsewhere it uses the two products of 0.7.0.
  • A decoding step's graph is kept between tokens.
  • Qwen3-ASR encodes up to four windows in one graph on a GPU.
  • On an Apple M5, Qwen3-ASR now decodes as fast as llama.cpp b11246 (0.6B: 146 to 158 tokens a second against 147 to 157).
  • On an RTX 2080 with Vulkan it recognizes the same clips 13 to 25% faster than 0.7.0.
  • Qwen3-TTS reaches its first audio and finishes sooner.
• Long audio uses less memory: 2.03 GB at the peak for 1338 s with Qwen3-ASR 0.6B, against 2.52 GB.
• Nothing a caller adapts to changed: the C API, the worker protocol, the GGUF layout and the files on Hugging Face are those of 0.7.0.

What's Changed

• Decode Qwen3 as fast as llama.cpp on a GPU, with flash attention where the device computes it by @nyosegawa in #49
• Raise the version to 0.7.1 by @nyosegawa in #50

Full Changelog: v0.7.0...v0.7.1