Repository navigation
Releases
Buddy 1.6.0
-
Windows: Buddy's quick mind has been to school. It practised on 30,000 simulated requests where the right answer was known, plus several hundred written by a language model playing different people, and learned how sure it can be of each kind of command. Before, that was a hand count from about 50 examples.
-
It now understands more ways of saying a simple command, 40 in all: "quit outlook", "bring up excel", "make notepad full screen", "put notepad away". Polite words no longer send a plain command to the slow brain: "can you close notepad" and "minimize excel please" are done at once.
-
It is more careful in two places. If you say "open edge" while Edge is already open, Buddy cannot know whether you want that window or a new one, so the brain decides. "It" and "that" ("close it") always go to the brain. "Close this" still does too.
-
Saying "no" over Buddy's reply in talk mode now counts against what Buddy just did at once, the same as typing it. "Stop" and "wait" only stop the voice.
-
How it was checked: on requests the simulator did not write (simulated users played by a second language model, Microsoft's Voice Access command list, our older test list and 5,500 requests written for other assistants), Buddy acted at once 114 times and never wrongly. When it said it was 98% sure, it was right every time; over all 233 readings its stated confidence was off by 0.028 on average.
-
Known limits: nearly all of this is simulated. No real person's requests and no real microphone have been measured yet. "Never wrong in 114" still allows an error rate of up to 2.6%. Buddy did only 42 of the simulated users' 191 one-step commands at once; the brain got the rest. On our older test list it does 40 of 49 at once, three fewer than 1.5.0, because of the "already open" caution. After you say "no" to a quick window action, closing a window by name goes back to the brain for a long time.
-
To switch it off: Settings, Voice input, untick "Do simple commands at once".
-
Mac: no changes in this version, only the version number.
Buddy 1.5.0
-
Windows: simple one-step commands are now done at once, without waiting for the brain. "Minimize excel", "switch to outlook", "open notepad" and "open google.com" no longer take a screenshot and a few seconds of thinking. In 15 real window actions on our test laptop this took 0.014 seconds (the slowest 0.092), where the free brain needed about 3.4 seconds just to reply.
-
Anything less plain is still thought over first: a question, two steps ("and", "then"), a "don't", an "if", "all" or "every", "could you" or "maybe", more than 8 words, highlighted text, or a window Buddy cannot pin down (two Notepads, or an app that is not open).
-
The more a mistake would cost, the surer Buddy has to be before it acts alone. That is why "close this" still goes to the brain.
-
Buddy picks up habits: when the brain does the same single step for the same short phrase again and again, Buddy starts doing it at once. Say or type "no", "not that", "wrong" or "undo" within 30 seconds and it counts against what Buddy just did, so the brain is asked again next time. Habits stay on your PC.
-
On 94 test requests the rules had never met, Buddy did 43 of the 49 one-step commands at once and none of them wrongly. Known limits: both test lists are typed English written by the same person who wrote the rules, 44 fast answers with no mistake still allows an error rate of up to about 7 %, a "no" said while Buddy is speaking in talk mode is not counted, and the habit numbers are not yet checked against real use.
-
To switch it off: Settings, Voice input, untick "Do simple commands at once".
-
Mac: no changes in this version, only the version number.
Buddy 1.4.0
-
Windows: talk mode. Press Ctrl+Alt+T, or pick "Talk with Buddy" in the tray menu, and have a spoken conversation without holding a key. Press it again to stop.
-
Buddy decides you have finished from your words, not only from a pause. On 192 test clips it was right about 89 to 93 % of the time, against 73 % for a fixed pause. The clips are computer-made voices we built ourselves, not recordings of real people.
-
You can interrupt Buddy while it is speaking: start talking and it stops and listens.
-
Known limits: on a busy laptop Buddy can take more than a second to notice you have finished (our 0.3 second target is not met yet). There is no echo cancelling, so use a headset in talk mode, or Buddy may hear its own voice.
-
Mac: no changes in this version, only the version number.
Buddy 1.3.2
-
Buddy for Mac is no longer English-only. Menu bar, Voice language: hold-to-talk listens in Indonesian, Chinese (Taiwan or Mainland), Cantonese, Japanese, Korean, Malay, Thai, Vietnamese, Hindi, German, French, Spanish, Portuguese or English.
-
On the Mac, Buddy answers in the language you asked in and reads the answer with a Mac voice for that language. On a test Mac, Apple's older language detector read Indonesian as Turkish; Buddy now uses the newer recogniser and your voice language as a hint.
-
On the Mac, menu items a non-English answer puts on the screen's top edge now snap to the matching text on screen, the Mac form of the 1.3.1 Windows fix.
-
Windows: the Focus tab's summary uses the whole width of the tab and says 1/5 stars instead of 1 stars.
Buddy 1.3.1
Buddy points right when it answers in your language.
- 1.3.0 began answering in the language of the question. A live test showed the cost: after a non-English answer the arrow often went to the top-left corner of the screen instead of the menu item (in 14 of 26 answers, the tour ended somewhere other than the item asked about).
- 1.3.1 ignores those corner points and takes a second look: it finds the named control among the controls Windows lists, otherwise it asks the AI once more, in English, only where the control is. In the same test, every one of 13 answers then ended on the right item (4 needed the second look).
- Every answer in that test came in the language of the question (26 of 26 with the 1.3.0 prompt; 1.2.0 answered Malay, Spanish and French in English).
- With Ollama on the processor only, Buddy skips the second look (about a minute) and shows no pointer rather than a wrong one.
Details and raw replies: bench/README.md, section "Reply language on a live brain".
Buddy 1.3.0
Buddy 1.3.0 is about voice input: 15 languages instead of 6, long dictations that no longer lose their end, and text that arrives sooner on a busy PC. The numbers below come from human recordings (FLEURS), not from synthetic voices.
What's new
- 15 language choices for voice input (was 6). New: Indonesian, Malay, Thai, Vietnamese, Chinese (Traditional) and Auto-detect; German, Spanish and French got their own entries.
- Each language now uses the engine that measured best on human recordings. Same clips, 1.2.0 -> 1.3.0: Indonesian 95.7% -> 29.1% word errors, Japanese 80.7% -> 22.1% and Korean 75.6% -> 13.8% character errors.
- Long dictations are decoded in parts while you hold the key. On a busy PC, a dictation of one to one and a half minutes in English, Mandarin or Spanish was typed 0.2 to 1.4 s after release instead of 13 to 20 s.
- English no longer loses what you say after 30 seconds: 9 wrong words of 98 in a 56-second recording, where 1.2.0 got 51 wrong.
- Short dictations type sooner on a busy PC: about 1.0 s after release instead of 3.0 to 4.8 s, with the same text.
- My words no longer rewrites everyday words or people's names: 0 wrong replacements on 721 test texts (1.2.0 made 56 on 684), 37,809 names from 63 languages are left alone, and a badly misheard term such as "Bizaira's" still becomes ZIRIS.
- Fillers are removed in 11 languages, spoken corrections work in 7 ("Rapatnya jam tiga, eh salah, jam empat" types "Rapatnya jam empat") and edit commands in 9. Buddy's AI is asked to answer in the language of the question.
- The first-use download shows size, speed and time left, unpacks 38 to 43% faster, and Buddy says how far it is if you try voice input meanwhile. Voice input has its own Settings tab.
Measured on human recordings
FLEURS clips, the same clips for both versions. Error means wrong words, or wrong characters for Chinese, Japanese, Korean and Thai.
| Language | Clips | Buddy 1.2.0 | Buddy 1.3.0 | Engine now |
|---|---|---|---|---|
| English | 20 | 12.7% | 12.7% | Whisper base.en |
| Mandarin | 20 | 10.3% | 10.3% | SenseVoice |
| Cantonese | 10 | 6.3% | 6.3% | SenseVoice |
| Indonesian | 20 | 95.7% | 29.1% | Omnilingual |
| Malay | 10 | 94.2% | 28.1% | Dolphin |
| Japanese | 10 | 80.7% | 22.1% | Dolphin |
| Korean | 10 | 75.6% | 13.8% | Dolphin |
| Thai | 10 | 99.8% | 19.2% | Dolphin |
| Vietnamese | 10 | 99.7% | 26.2% | Dolphin |
| German | 10 | 18.5% | 18.5% | Parakeet V3 |
| Spanish | 10 | 1.4% | 1.4% | Parakeet V3 |
| French | 10 | 8.2% | 8.2% | Parakeet V3 |
No engine typed anything on the silence clips. The scripts, the result files and every table are in bench/README.md.
Limits
- No calm PC was measured: every timing above was taken at 87 to 100% processor load, because other programs were running on the test PC. A second run, taken while another program held the PC at 100%, repeated the result: 18.40 s to 1.50 s as it was and 25.78 s to 1.71 s with every core busy, shipped faster and same text on 6 of 6 clips in both rows.
- Indonesian is still at 29.1% word errors. Auto-detect is less accurate than choosing the language (Indonesian 54.8% against 29.8% on the same clips) and has the 30-second limit.
- Speech much quieter than what came before it can be taken for background and left out. 1.2.0 has the same limit at the start and end of a dictation.
- Japanese, Korean, Thai and Vietnamese get filler removal only, no spoken corrections.
- Natural reading voices exist for English, Chinese, Spanish, French, Hindi, Italian and Portuguese. Other answers are read by the chosen voice.
- The reply-language rule was not tried against a live model, and the new settings layout was checked by measured window sizes, not seen on screen.
Download
Windows: Buddy-win64.zip (unzip, run Buddy.exe). This release has the Windows build only; the Mac and iOS downloads stay at 1.2.0.
Buddy 1.2.0 - type by voice, offline (beats Windows speech and Handy's engine)
Buddy can now type what you say into any app, fully offline, and it's more accurate than both Handy's default engine and the Windows recognizer it used before.
New: voice dictation (Ctrl+Alt+D)
- Hold Ctrl+Alt+D, talk, let go: the text lands in mail, chat, PLM forms or Word. Tap once for hands-free, tap again to type. Esc cancels. Never touches the clipboard.
- Drops "um"/"uh", follows spoken corrections ("at three, no wait, at four" types "at four"; "... scratch that" drops the sentence).
- My words: your jargon, names and part numbers get fixed after recognition. Snippets expand short phrases.
- Select text first and say "make this shorter" to edit it by voice.
- Settings -> Type last dictation again / Copy last dictation; history of the last 20.
New speech engines (Settings -> Voice input -> Speech model: Auto)
- English: Whisper base.en + Silero voice detection. Chinese / Cantonese / Japanese / Korean: SenseVoice. 24 other European languages: Parakeet V3.
- Models (~170-490 MB) download once from GitHub in the background (works where Hugging Face is blocked), resume after network drops; the Windows recognizer answers until they're ready.
Measured (24 synthetic clips, clean + office noise; scripts and audio attached):
| English WER clean / noisy | Made-up text on 4 silent clips | |
|---|---|---|
| Windows recognizer (Buddy 1.1) | 37.4% / 86.9% | 1 |
| Parakeet V3 + VAD (Handy's recommended engine) | 13.6% / 15.4% | 0 |
| Buddy 1.2 (Whisper base.en + VAD + My words) | 4.9% / 4.9% | 0 |
Chinese character error rate: 42% / 91% -> 0.0% / 1.5%. Warm latency 0.88 s on a calm laptop CPU (2.1 s when the CPU was saturated). Wispr Flow is cloud-only, so it's compared on published features in the README, not measured.
Full comparison: https://github.com/juw12/buddy#voice-buddy-vs-wispr-flow-vs-handy
buddy-voice-bench-audio.zip = the benchmark clips, so anyone can re-run bench/.
Buddy 1.1.3 - tidy updates
Buddy 1.1.3
- Tidies half-finished update downloads every time Buddy starts
- Removes the old-version safety copy once the new Buddy has run fine for 10 minutes
- Includes 1.1.2: the hidden update helper now starts reliably on every PC, and the island shows download progress
Buddy 1.1.1 - safer auto-update
A small safety fix, delivered by Buddy's own auto-update.
- Safer updates on Windows: if the update helper runs again after an update is already installed, it now leaves Buddy alone instead of swapping folders and rolling back.
- Mac and iPhone keep their updates when a release only ships a Windows fix (the update manifest carries them forward from the last release that has them).
Windows only: Buddy-win64.zip. Mac and iPhone stay on 1.1.0, which is still the latest for them.
Buddy 1.1 - iPhone Dynamic Island + auto-update
Buddy 1.1 - Buddy on your iPhone's Dynamic Island, and Buddy now updates itself.
- iPhone app (new): Buddy lives in the real Dynamic Island as a Live Activity, and on the Lock Screen. Set Back Tap or the Action Button to "Ask Buddy about my screen" and the answer appears in the island without leaving your app. Also: Siri ("Ask Buddy"), a Home/Lock Screen widget, read-aloud. Free sideload: install
Buddy-iOS.ipawith Sideloadly or AltStore and your Apple ID (guide). - Auto-update on Windows and Mac: Buddy checks for a new version a minute after it starts and every 3 hours. It downloads it, checks its SHA-256, self-tests it, and swaps only while you're idle. Your settings, voices and history stay; if anything fails, the old version comes back. Tray / menu bar: Check for updates.
- iPhone updates: the app shows a banner when a new version is out. AltStore users can add the source
https://juw12.github.io/buddy/altstore.json. - Web island: refreshes itself to the newest version.
Downloads: Buddy-win64.zip (Windows 10/11), Buddy-mac.zip (macOS), Buddy-iOS.ipa (iPhone, iOS 17+, sideload).
Honest limits: the iPhone app is unsigned, so it has to be signed with your own Apple ID (free accounts re-sign every 7 days; AltStore does it for you). It's tested in CI on the iOS Simulator, not yet on a physical iPhone.