Repository navigation
Release 0.4.5: local Qwen3 (MLX) engine, settings redesign, refine prompt, Soniox v5 - #54
Merged
Merged
Conversation
Move realtime/async to stt-rt-v5 / stt-async-v5 and disable automatic
endpoint detection: a long thinking pause could bake a sentence mark into
the middle of a phrase ("一种 [13s pause] 英雄式" → "一种。英雄式"). The
user's stop action already flushes via the end-of-audio marker. Adds a
quote-pairing regression test for a sentence mark inside a quotation.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Settings gains a 润色提示词 editor (save / restore default); a custom prompt fully replaces the default rules and is used by both dictation and the connection test. TextRefiner logs a per-request trace (outcome, lengths, finish reason, timing) and records whether the clipboard changed while a deferred refinement was in flight. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Release 0.4.5.
MLXASRengine target built on mlx-audio-swift, pinned to244458d. Streaming output drives the HUD as a live preview. On stop, the whole recording is transcribed once more and that result is what gets pasted; streaming text is the fallback. The full pass uses a repetition penalty and gets glossary context. Only real-time streaming models are supported.ingestjust appends under a lock, so it is safe on the audio thread.Qwen3; the storedlocalid migrates automatically.make buildnow usesxcodebuild, which compiles MLX's Metal shaders. Resource bundles go intoContents/Resources.Package.resolvedis committed.stt-rt-v5/stt-async-v5models. Endpoint detection is off, because a long pause was inserting a sentence break mid-phrase.smoke-cli mlx-transcribefor headless real-model checks; it can switch the final pass off and pass context.Verification
swift teston the merged branch: 430 tests, 0 failures.smoke-clion the author's recordings: 5 clips × {0.6B, 1.7B} × {zh-CN, auto}, all using the full-pass result.make installand tested by hand in the app: live preview, final paste, switching to 1.7B, and the settings cards.Known follow-ups
🤖 Generated with Claude Code