github kizuna-ai-lab/sokuji v0.15.12

latest release: v0.15.13
17 hours ago

What's Changed

Features

  • WebGPU shader-f16 dtype selection: Variant-aware model system with automatic GPU capability detection, upgrade hints, and incompatibility warnings
  • Multilingual translation models: Language priority sorting and multilingual translation model support
  • Local inference improvements: Add local.asr.start event for speech detection; validate TTS model language compatibility before session start
  • Model management: Add clear-all button to delete all downloaded models
  • Logging: Add modelId to ASR, translation, and TTS log events; include systemPrompt in translation result and log events
  • Parakeet TDT: Add 8 EU language display names for Parakeet TDT model

Bug Fixes

  • Fix Whisper WebGPU hallucination by switching to Xenova repos and q4 dtype
  • Populate WebGPU cache before model readiness checks on init
  • Fix streaming ASR: pass model type to createOnlineRecognizer
  • Only use /no_think for Qwen3+ models, not Qwen2.5
  • Fix blob URL leak and localStorage safety
  • Fix accessibility: add tabpanel role to LogsPanel content area

UI/UX

  • Align LogsPanel and Settings header/tab-bar heights and font sizes
  • Use shared TabBar component in LogsPanel for consistent tab styling

Other

  • Remove Qwen 2.5 0.5B translation model from manifest
  • Remove multilingual flag from Parakeet TDT manifest entry

Don't miss a new sokuji release

NewReleases is sending notifications on new releases.