Skip to content

CHANGELOG.md

Latest commit

 

History

History
80 lines (66 loc) · 3.56 KB

CHANGELOG.md

File metadata and controls

80 lines (66 loc) · 3.56 KB
Raw

Changelog

1.7.1 (2026-05-01)

Rebranding — plugin and skill renamed to reflect multi-engine support

Since 1.7.0 added MiMo-V2.5-ASR alongside FunASR, the plugin is no longer FunASR-only. This release renames repo, plugin, skill, and script paths accordingly. This affects users who invoke the CLI by path or import the script as a Python module; does not affect users on ClawHub or skills.sh.

Renames:

  • GitHub repo: zxkane/audio-transcriber-funasr → zxkane/audio-transcriber (old URLs continue to work via GitHub's automatic redirect)
  • Plugin dir: plugins/funasr-transcriber/ → plugins/audio-transcriber/
  • Skill dir: skills/funasr-transcribe/ → skills/audio-transcribe/
  • Skill name (SKILL.md name:): funasr-transcribe → audio-transcribe
  • Main CLI script: scripts/transcribe_funasr.py → scripts/transcribe.py
  • .claude/skills/funasr-transcribe symlink → .claude/skills/audio-transcribe

Not renamed (intentional, backward compat):

  • ClawHub package slug zxkane-audio-transcriber-funasr is preserved — existing installs continue to work, no migration needed for /plugin install audio-transcriber@zxkane-audio-transcriber-funasr.
  • Internal function transcribe_with_funasr() keeps its name (it is specifically the FunASR-backed path, so the name is accurate).

Migration for direct-invocation users

If you were invoking the script by path, update your command:

- python3 plugins/funasr-transcriber/skills/funasr-transcribe/scripts/transcribe_funasr.py
+ python3 plugins/audio-transcriber/skills/audio-transcribe/scripts/transcribe.py

If you were importing as a Python module (rare, mostly test-only):

- import transcribe_funasr
+ import transcribe

1.7.0 (2026-04-29)

Added

  • --lang mimo: local inference with Xiaomi's MiMo-V2.5-ASR (8B, GPU-only), reusing FSMN VAD + CAM++ diarization so output format matches --lang zh.
  • scripts/setup_mimo.sh: opt-in installer (INSTALL_MIMO=1 bash setup_env.sh) that clones the MiMo repo, installs flash-attn, and downloads ~20 GB of weights to $MIMO_WEIGHTS_PATH (defaults to $HF_HOME → ~/.cache/huggingface). Prefers pre-built flash-attn wheels matching the installed torch + Python + ABI so the install completes on CUDA-driver-only hosts (most AWS GPU instances) without requiring the CUDA toolkit (nvcc). Falls back to a source build only when no matching wheel is published. The MiMo GitHub repo is pinned to a known-good commit (validated against our MimoAudio.__init__ kwarg contract and runtime deps); override with MIMO_PINNED_COMMIT=<sha> to trial a newer upstream.
  • --mimo-audio-tag, --mimo-batch, --mimo-weights-path, --resume-mimo: CLI flags for the new preset. --resume-mimo picks up from a mid-run failure using *_mimo_partial.json with audio-hash verification.
  • New tests in scripts/test_mimo_asr.py (mocked; GPU-free, CI-safe).

Changed — BREAKING ENVIRONMENT CHANGE

  • setup_env.sh now requires Python 3.12. Existing .venv/ directories created with earlier Python versions are detected and rebuilt on the next run of setup_env.sh. Expect a 2–3 GB re-download of FunASR dependencies.
  • New dependencies: scikit-learn (for KMeans clustering in the MiMo path), soundfile (already transitively present, now explicit).

Notes

  • --lang mimo hard-fails if CUDA is unavailable or VRAM < 20 GB. There is no CPU fallback for this preset.
  • --hotwords and --batch-size are silently ignored with --lang mimo; use --mimo-batch for per-call concurrency.