The slowest moment in voice input is after you stop talking

The SGLang Omni team releases Voxt 2.0. A local voice workflow for macOS that puts voice input, translation, rewriting, and answer cards into a menu bar app. The team put it plainly: rather than keep a productivity tool that works well for themselves, they would open source it and let the community use it too.

First, the demo:

The bottleneck in voice input is often the waiting period after you stop. Supporting ten more languages is nice, but whether users will use it every day depends first on whether they have to go pour a glass of water after they stop recording.

Voxt 2.0 was refactored from Swift to Go, and during long recordings it recognizes the parts already spoken while you are still talking. Official numbers: on an M4 Mac with 24GB, for a 4 minute 46 second dictation, the wait after stopping drops from 32.3 seconds to 2.6 seconds; for very long input, the wait is 92% shorter than in 1.x. That number is more likely to affect usage frequency than a few extra languages.

The team says even a five-year-old M1 Mac can run Qwen3-ASR locally, then hand polishing off to a remote LLM API. In a reply, a commenter named Mr. Grayscale said this division of labor is right: recognition needs to be fast and private, while only polishing needs a large model.

On the feature side, 2.0 keeps and expands three main paths:

• Voice input: triggered by Right Option; press it again when you finish speaking, and the text lands at the cursor. After ASR recognition, an LLM can remove filler words, add punctuation, and fix self-corrections.
• Translation: Right Option plus Shift; select text or just speak, and hand it to an LLM to translate into the target language. The README lists 22 target languages.
• Rewrite and assistant: Control plus Right Option; select any text and say 'make it shorter,' 'make it more formal,' or 'explain it.' The result goes into an answer card, where you can continue asking follow-up questions by voice, then copy or insert it.

2.0 also adds several practical details: if recognition fails, the recording is kept and can be retried from history; answer cards support continuous interaction; dictionaries, history, and compatible speech models from Voxt 1.x can be migrated in one click; API keys and shortcuts are not imported and must be set again. New features include ChatGPT login, fallback translation models, custom prompts, and Markdown and JSONL history export.

The model ecosystem is another focus for Voxt. By default, speech runs on the Mac via the SGLang Omni MLX runtime, native C++, not Python. The recommended model, Qwen3-ASR 1.7B 8bit, is about 2.5GB; there is also a 0.6B 4bit model at about 713MB. It also supports MOSS-Transcribe-Diarize and cloud ASR APIs such as OpenAI Transcribe, Doubao, Zhipu GLM-ASR, Alibaba Bailian, StepFun, Xiaomi MiMo, and Google Gemini.

On the LLM side, it connects to any OpenAI-compatible API: locally you can use SGLang MLX, Ollama, LM Studio, or oMLX; remotely you can use DeepSeek V4.1 or Kimi K3.1, which the team says they use often. Where data goes is up to the user. By default, speech stays on the Mac, only text is sent to the configured LLM, API keys are stored in the macOS Keychain, and there are no accounts and no analytics.

Qwen3-ASR can also be accelerated with DFlash transcription: download a small draft model and use speculative decoding to make the speech model transcribe faster; removing it restores the original setup.

The installation requirements are clear: Apple silicon, macOS 15 or later. Intel Macs should continue using Voxt 1.x. The app is about 280MB, and the recommended speech model is 2.5GB. Voxt 2 does not auto-update; you can install over it with the dmg, or use Homebrew:

brew install --cask hehehai/tap/voxt
brew upgrade --cask voxt

Download dmg: https://github.com/sgl-project/sglang-omni/releases/download/voxt-v2.0.3/Voxt-2.0.3-arm64.dmg

Project README: https://github.com/sgl-project/sglang-omni/blob/main/Voxt/README.md

Website: https://voxt.actnow.dev/

Chinese demo video: https://www.youtube.com/watch?v=mOwWavTSXu0

The official pricing page says local use is free forever, remote models use your own key, and there is an optional $8 one-time support option. The code is open source under Apache License 2.0. The website also lists notes, meeting transcription, and recording analysis, but 2.0's main focus is still the three things: input, translation, and rewriting.