Skip to content

PandaCode 0.3.108

Choose a tag to compare

@github-actions github-actions released this 10 Nov 14:13

Voice dictation with on-device speech models

New

  • Dictate into the composer. Press the microphone button, or ⌘/Ctrl+Shift+D while focus is in the prompt box, on the microphone button or on the dictation status card, and the prompt box becomes a live, scrolling waveform while you speak. Press Stop or Enter and the box is back at once, so you can keep typing while the speech is transcribed; the text lands where you were typing. Escape cancels a recording, and discards a transcript waiting in review when focus is inside the review card. The shortcut can toggle, hold to record, or do both.
  • Dictation never sends automatically: review the draft and send it yourself. When a recording reaches two minutes or 16 MiB it stops; press Transcribe recording to transcribe it. If nothing is set up yet, the first microphone click opens Settings → Voice.
  • Failed transcriptions can be retried. If a transcription fails, the recording is kept so you can retry it. It is dropped when you discard it, choose Record again, delete the thread the dictation started in, or quit PandaCode. If you edit the draft heavily while it is being transcribed, the text is offered for review instead of being dropped in the wrong place. Switching threads stops the recording and delivers the text to the thread where you started dictating: it is inserted when you go back, or held there for review. Closing the window while PandaCode keeps running also stops the recording and transcribes it. When closing the window quits PandaCode (always on Linux; on macOS and Windows only if the menu-bar or tray icon couldn't be created), or when you quit the app, a dictation that is still recording is cancelled and its audio isn't uploaded, and a transcription still in progress is lost; audio already sent to a speech service for transcription can't be recalled.
  • Settings → Voice. A new page with three cards:
    • Input: choose a microphone, test it with a live level meter and playback, and pick the shortcut behaviour.
    • Transcription: send audio to an on-device model, a speech service, or turn voice off, and choose a language. The model browser sits here: search languages and models, see each model's accuracy grade and error rate per language, speed, size and licence, then download, test, select or delete it.
    • Speech services: presets for OpenAI, Groq and Mistral with your own key (kept in PandaCode's secret store), or Local or custom server: a full transcription URL, such as a whisper.cpp server; the key is optional. PandaCode sends the audio as a WAV file in a multipart upload (file, model and, if set, language fields), and the server must reply with JSON containing text; chat or text-generation endpoints won't work. Remote servers must use HTTPS, plain HTTP works only with a literal loopback address such as 127.0.0.1 or [::1] (not the name localhost), and redirects are refused. The URL must have a path and no query string, credentials or fragment; a key, if set, is sent as Authorization: Bearer <key>.
  • On-device speech. Models run on your computer with transcribe.cpp (Metal on Apple silicon, CPU elsewhere), so audio never leaves it. Downloads come from Hugging Face, resume after interruption and are checked against a SHA-256 checksum before use.

Changed

  • When a send fails, your draft is now always put back, ahead of anything you typed since. Before, it came back only if nothing new had been typed. This applies to every send, not only dictated text.
  • Settings now has nine sections; Voice sits between Appearance and Agents.
  • Settings → About has a Third-party notices button that opens the licences bundled with the app.

Privacy

  • Nothing is recorded until you start dictation or run a microphone or model test, and nothing is uploaded unless you choose a speech service. With one chosen, audio is uploaded when you stop or transcribe a recording (Stop, Enter, Transcribe recording, Retry) or when switching threads or anything else closes the prompt box mid-recording while PandaCode keeps running (such as closing the window to the menu bar or tray). Automatic stops (the two-minute or 16 MiB limit, or a failing microphone) upload nothing and keep the audio until you press Transcribe recording. Model downloads contact huggingface.co (and its *.huggingface.co / *.hf.co download hosts) only when you press Download. Both are listed in the README privacy table.

Known limitations

  • On macOS this release crashes when you open Settings → Voice, start dictation with the microphone button or ⌘/Ctrl+Shift+D, or press Test microphone. Update to 0.3.109, which fixes it.
  • On Windows and Linux x64, on-device models need a processor with SSE4.2, AVX and F16C (most made since about 2013; some low-end Pentium and Celeron chips lack AVX). On a processor without them the speech engine stops when a model loads, and the app shows “The local speech engine stopped unexpectedly” and suggests retrying, but on such a processor retrying won't help. PandaCode doesn't detect this in advance; use a speech service instead.
  • On-device models need a processor with the Armv8.2 dot-product and i8mm instructions on Windows on Arm; older Snapdragon 8cx laptops are told to use a speech service instead.
  • The Windows on Arm i8mm check uses a processor-feature id that only newer Windows 11 builds report. Older Windows 11 builds and Windows 10 on Arm don't, so on those even capable processors (such as Snapdragon X) are told they can't run on-device models. Updating Windows, or using a speech service, resolves it.
  • Voice uploads and model downloads ignore the system and HTTP proxy settings, so on a network that requires a proxy they fail.
  • On Linux the AppImage uses the system's ALSA library (libasound.so.2) for the microphone.
  • Tested in the browser preview (Chromium and WebKit) and with the native Rust test suite. The real speech engine has only been run on Apple silicon: its tests are ignored by default, and CI's Linux job skips them. It hasn't been run on Linux x64, Windows, Intel Mac or Linux on Arm. Installers for all six platforms built and are attached, but they haven't been tested by hand. A real microphone, the macOS permission prompt and real speech providers have not been tested by hand.

Installing

  • macOS (Apple silicon): open PandaCode-macOS-arm64.dmg and drag PandaCode to Applications. This build is signed and notarized, so macOS opens it without a warning.
  • macOS (Intel): the same, with PandaCode-macOS-x64.dmg.
  • Windows x64: run PandaCode-Windows-x64-setup.exe. If SmartScreen warns about an unknown publisher, choose More info > Run anyway.
  • Windows arm64: run PandaCode-Windows-arm64-setup.exe (same SmartScreen step).
  • Linux x64: chmod +x PandaCode-Linux-x86_64.AppImage and run it.
  • Linux arm64: chmod +x PandaCode-Linux-aarch64.AppImage and run it.
  • Any of the above, from a terminal: npx pandacode downloads, verifies and installs the right build, and signs it locally on macOS, so no Gatekeeper warning appears.

The installers each have a .sha256 checksum beside them. Installed copies update themselves from this page. Build from source: see CONTRIBUTING.md.