ZillaNote lite 0.4.2

Record a meeting, get minutes. ZillaNote lite records on your computer, finds who spoke, transcribes locally with Qwen3-ASR (or with a speech service you choose), writes the minutes with your ChatGPT plan or the language model you point it at, and mails them to you.

New in 0.4.2

• The bar lies down. The small bar at the edge of the screen is now horizontal, at the same size (142 by 38 points instead of 38 by 142): the logo, the record button, the level meter over the time, the progress ring while a meeting is processed, and the violet tab that opens the window at its right end. It still starts at the right edge of the screen and can be dragged anywhere by its logo.
• A bigger logo on the bar, the size of the record button.

Also in 0.4: Forget removes a voice's voice print from every meeting, voice recognition can be turned off (0.4.1); minutes on your ChatGPT plan, no needless speech-model download, and upgrades that keep everything you set up (0.4.0).

Install on a Mac

1. Apple silicon, macOS 12 or later (14.2 or later for the computer's sound and for noticing calls).
2. Open ZillaNote_0.4.2_aarch64.dmg and drag ZillaNote to Applications.
3. The app is not signed with an Apple Developer ID, so macOS blocks the first start of a downloaded copy: open System Settings → Privacy & Security, find "ZillaNote was blocked", press Open Anyway. Or in Terminal: xattr -dr com.apple.quarantine /Applications/ZillaNote.app
4. Allow the microphone, and "System Audio Recording" when the first recording asks.

Install on Windows

1. Windows 10 or 11, x64. Run ZillaNote_0.4.2_x64-setup.exe; it installs for the current user and needs no administrator.
2. The installer is not signed, so SmartScreen says "Windows protected your PC": choose More info, then Run anyway.
3. If Windows asks, allow the microphone for desktop apps (Settings → Privacy & security → Microphone).

Then, on either

The window offers to download the speech model and the speaker models (32 MB), unless a speech model is already on the computer; with a speech service chosen in Settings → Speech recognition, only the speaker models. In Settings → Language model, press Continue with ChatGPT, or enter a model of your own (any OpenAI-compatible endpoint: LM Studio, Ollama, or a hosted API with your key); under Email the minutes, add the mail account if you want them mailed.

Known limits

• Sign in with ChatGPT is a preview of OpenAI's: what a plan allows may change on OpenAI's side. A ChatGPT plan pays for the minutes only, not for transcription. It has been used with a real ChatGPT account on a Mac, not yet on Windows.
• The speech-service option has been tested against a stand-in service, not yet against OpenAI, Groq or SiliconFlow themselves.
• The call watch has been checked against a test program that opens the microphone, not yet through a whole Teams or Zoom call on either system.
• Windows is a preview: built, unit-tested and started on a Windows machine by the build job, installed and started on a real PC, not yet through a whole meeting there.
• Neither installer is signed by a registered developer (see the install steps). On the Mac the keychain asks again after every update, because each build has a new signature: choose Always Allow; nothing has to be typed again.
• On Windows the API keys, the mail password and the ChatGPT sign-in are kept in files of your user profile, not in the Credential Manager yet; only WAV recordings can be imported; transcription on the computer runs on the CPU.
• Outlook.com accounts cannot send the mail: Microsoft no longer accepts passwords over SMTP for them.

SHA-256 of ZillaNote_0.4.2_aarch64.dmg: 04e660ff7324b72db0072c4b6ca2ca834977acf49ce4e4b5d3f5cbb98a98ff14

SHA-256 of ZillaNote_0.4.2_x64-setup.exe: 622482cfa0199e83640f619090a72673244f29203801c417ecbd6323f5e3155b (built from this tag by the repository's build job, which also ran the unit tests on Windows, started the built app there and measured its bar)

The same two files are also attached as ZillaNote_aarch64.dmg and ZillaNote_x64-setup.exe, the names that the download buttons on the website and in the README always fetch from the latest release.

ZillaNote lite 0.4.1

Record a meeting, get minutes. ZillaNote lite records on your computer, finds who spoke, transcribes locally with Qwen3-ASR (or with a speech service you choose), writes the minutes with your ChatGPT plan or the language model you point it at, and mails them to you.

New in 0.4.1

• Forget means forget. Forgetting a remembered voice now also takes its voice print out of every meeting it was named or recognized in, not only out of the list. The name already shown in those transcripts stays, as text. Forget all voices does it for every voice at once.
• Voice recognition can be turned off. Settings → Remembered voices → "Recognize named voices in later meetings". Off, nobody is recognized, no voice print is kept for new meetings, and naming a speaker labels them in that meeting only. It stays on for everyone who has it now.
• Asking before naming. Where you type a speaker's name, ZillaNote now says that it will know the voice in later meetings, and to name only people who agreed.
• What is kept, and for how long. How it works now explains voice prints (256 numbers per voice, not a recording, never leaving the computer), where they are kept and until when, and the README has a Responsible use section.

Also in 0.4: minutes on your ChatGPT plan (Settings → Language model → Continue with ChatGPT), no needless speech-model download when one is already on the computer, and upgrades that keep everything you set up. See 0.4.0.

Install on a Mac

1. Apple silicon, macOS 12 or later (14.2 or later for the computer's sound and for noticing calls).
2. Open ZillaNote_0.4.1_aarch64.dmg and drag ZillaNote to Applications.
3. The app is not signed with an Apple Developer ID, so macOS blocks the first start of a downloaded copy: open System Settings → Privacy & Security, find "ZillaNote was blocked", press Open Anyway. Or in Terminal: xattr -dr com.apple.quarantine /Applications/ZillaNote.app
4. Allow the microphone, and "System Audio Recording" when the first recording asks.

Install on Windows

1. Windows 10 or 11, x64. Run ZillaNote_0.4.1_x64-setup.exe; it installs for the current user and needs no administrator.
2. The installer is not signed, so SmartScreen says "Windows protected your PC": choose More info, then Run anyway.
3. If Windows asks, allow the microphone for desktop apps (Settings → Privacy & security → Microphone).

Then, on either

The window offers to download the speech model and the speaker models (32 MB), unless a speech model is already on the computer; with a speech service chosen in Settings → Speech recognition, only the speaker models. In Settings → Language model, press Continue with ChatGPT, or enter a model of your own (any OpenAI-compatible endpoint: LM Studio, Ollama, or a hosted API with your key); under Email the minutes, add the mail account if you want them mailed.

Known limits

• Sign in with ChatGPT is a preview of OpenAI's: what a plan allows may change on OpenAI's side. A ChatGPT plan pays for the minutes only, not for transcription. It has been used with a real ChatGPT account on a Mac, not yet on Windows.
• The speech-service option has been tested against a stand-in service, not yet against OpenAI, Groq or SiliconFlow themselves.
• The call watch has been checked against a test program that opens the microphone, not yet through a whole Teams or Zoom call on either system.
• Windows is a preview: built, unit-tested and started on a Windows machine by the build job, installed and started on a real PC, not yet through a whole meeting there.
• Neither installer is signed by a registered developer (see the install steps). On the Mac the keychain asks again after every update, because each build has a new signature: choose Always Allow; nothing has to be typed again.
• On Windows the API keys, the mail password and the ChatGPT sign-in are kept in files of your user profile, not in the Credential Manager yet; only WAV recordings can be imported; transcription on the computer runs on the CPU.
• Outlook.com accounts cannot send the mail: Microsoft no longer accepts passwords over SMTP for them.

SHA-256 of ZillaNote_0.4.1_aarch64.dmg: 14ec3d8f521abba4e63788ca7db4aae40c90ae1545a40e173fac53087335a5b4

SHA-256 of ZillaNote_0.4.1_x64-setup.exe: 954b0e201a2b9ec919dc9a9dff148c287762611206b7ae0fdfffcb0e357aa614 (built from this tag by the repository's build job, which also ran the unit tests on Windows, started the built app there and measured its bar)

The same two files are also attached as ZillaNote_aarch64.dmg and ZillaNote_x64-setup.exe, the names that the download buttons on the website and in the README always fetch from the latest release.

ZillaNote lite 0.4.0

Record a meeting, get minutes. ZillaNote lite records on your computer, finds who spoke, transcribes locally with Qwen3-ASR (or with a speech service you choose), writes the minutes with your ChatGPT plan or the language model you point it at, and mails them to you.

New in 0.4.0

• Minutes on your ChatGPT plan. Settings → Language model → Continue with ChatGPT: sign in once in your browser, pick one of the models your account offers, and the minutes are written on your ChatGPT plan, with no API key and no cost beyond your plan. This uses OpenAI's Sign in with ChatGPT for open-source apps. The transcript is sent to OpenAI under your account, with nothing stored there by the request (store: false). Your plan's limit for ZillaNote can be seen and set under Manage usage (chatgpt.com/settings/usage); when it is reached, the meeting keeps its transcript and says so, and Write minutes tries again later. Your own model (LM Studio, Ollama, or an API with a key) stays the default and works as before.
• No needless speech-model download. Until you pick a speech model yourself, ZillaNote uses one that is already on your computer (left by an earlier version, or in LM Studio's folder) rather than asking you to download the 0.6B default.
• Upgrades keep everything you set up. Your settings, system prompt, vocabulary, email setup and remembered voices come through the update unchanged; this is now checked by tests on every release. A settings or voices file that cannot be read is set aside (*.unreadable-<time>.json) and never saved over.

Install on a Mac

1. Apple silicon, macOS 12 or later (14.2 or later for the computer's sound and for noticing calls).
2. Open ZillaNote_0.4.0_aarch64.dmg and drag ZillaNote to Applications.
3. The app is not signed with an Apple Developer ID, so macOS blocks the first start of a downloaded copy: open System Settings → Privacy & Security, find "ZillaNote was blocked", press Open Anyway. Or in Terminal: xattr -dr com.apple.quarantine /Applications/ZillaNote.app
4. Allow the microphone, and "System Audio Recording" when the first recording asks.

Install on Windows

1. Windows 10 or 11, x64. Run ZillaNote_0.4.0_x64-setup.exe; it installs for the current user and needs no administrator.
2. The installer is not signed, so SmartScreen says "Windows protected your PC": choose More info, then Run anyway.
3. If Windows asks, allow the microphone for desktop apps (Settings → Privacy & security → Microphone).

Then, on either

The window offers to download the speech model and the speaker models (32 MB), unless a speech model is already on the computer; with a speech service chosen in Settings → Speech recognition, only the speaker models. In Settings → Language model, press Continue with ChatGPT, or enter a model of your own (any OpenAI-compatible endpoint: LM Studio, Ollama, or a hosted API with your key); under Email the minutes, add the mail account if you want them mailed.

Known limits

• Sign in with ChatGPT is a preview of OpenAI's: what a plan allows may change on OpenAI's side. A ChatGPT plan pays for the minutes only, not for transcription. It has been used with a real ChatGPT account on a Mac, not yet on Windows.
• The speech-service option has been tested against a stand-in service, not yet against OpenAI, Groq or SiliconFlow themselves.
• The call watch has been checked against a test program that opens the microphone, not yet through a whole Teams or Zoom call on either system.
• Windows is a preview: built, unit-tested and started on a Windows machine by the build job, installed and started on a real PC, not yet through a whole meeting there.
• Neither installer is signed by a registered developer (see the install steps). On the Mac the keychain asks again after every update, because each build has a new signature: choose Always Allow; nothing has to be typed again.
• On Windows the API keys, the mail password and the ChatGPT sign-in are kept in files of your user profile, not in the Credential Manager yet; only WAV recordings can be imported; transcription on the computer runs on the CPU.
• Outlook.com accounts cannot send the mail: Microsoft no longer accepts passwords over SMTP for them.

SHA-256 of ZillaNote_0.4.0_aarch64.dmg: cd310646d5eff578a9a2e78c60103c11c4dd65cb86f907c7d79a4230476ec0ea

SHA-256 of ZillaNote_0.4.0_x64-setup.exe: 8f9983673ee1a9aa24edf1d033f6cf3c63449140865bdbf6dd6099c4856e7898 (built from this tag by the repository's build job, which also ran the unit tests on Windows, started the built app there and measured its bar)

The same two files are also attached as ZillaNote_aarch64.dmg and ZillaNote_x64-setup.exe, the names that the download buttons on the website and in the README always fetch from the latest release.

ZillaNote lite 0.3.1

Record a meeting, get minutes. ZillaNote lite records on your computer, finds who spoke, transcribes locally with Qwen3-ASR (or with a speech service you choose), writes the minutes with the language model you point it at, and mails them to you.

New in 0.3.1

• Settings in two levels. The first page keeps only the switches for recording and minutes. Speech recognition, the language model, prompt and vocabulary, email and remembered voices each have a page of their own, behind a row that says what is set there, or what is missing.
• Your own speech service. Speech recognition can now be a service with OpenAI's transcription API (OpenAI, Groq, SiliconFlow, a Whisper server of your own) instead of Qwen3-ASR on your computer: endpoint, model and API key, with a button to test it. Speakers are still found on your computer, and no speech model has to be downloaded. With a service chosen, each stretch of speech is sent to it; transcribing on your computer stays the default.
• Delete a speech model. A downloaded model has a Delete button, to give the space back after switching to the other model or to a service. Only ZillaNote's own copy is deleted, never files LM Studio keeps.
• How to name a voice is on the screen. Settings → Remembered voices and the speaker chips above a transcript now say how to name a speaker, so the voice is known in later meetings.
• The Settings gear is as big as the icons next to it.

Install on a Mac

1. Apple silicon, macOS 12 or later (14.2 or later for the computer's sound and for noticing calls).
2. Open ZillaNote_0.3.1_aarch64.dmg and drag ZillaNote to Applications.
3. The app is not signed with an Apple Developer ID, so macOS blocks the first start of a downloaded copy: open System Settings → Privacy & Security, find "ZillaNote was blocked", press Open Anyway. Or in Terminal: xattr -dr com.apple.quarantine /Applications/ZillaNote.app
4. Allow the microphone, and "System Audio Recording" when the first recording asks.

Install on Windows

1. Windows 10 or 11, x64. Run ZillaNote_0.3.1_x64-setup.exe; it installs for the current user and needs no administrator.
2. The installer is not signed, so SmartScreen says "Windows protected your PC": choose More info, then Run anyway.
3. If Windows asks, allow the microphone for desktop apps (Settings → Privacy & security → Microphone).

Then, on either

The window offers to download the speech model and the speaker models (32 MB); with a speech service chosen in Settings → Speech recognition, only the speaker models. In Settings → Language model, enter the model for the minutes (any OpenAI-compatible endpoint: LM Studio, Ollama, or a hosted API with your key) and, under Email the minutes, the mail account if you want them mailed.

Known limits

• The speech-service option has been tested against a stand-in service, not yet against OpenAI, Groq or SiliconFlow themselves. Please report what you see.
• The call watch has been checked against a test program that opens the microphone, not yet through a whole Teams or Zoom call on either system.
• Windows is a preview: built, unit-tested and started on a Windows machine by the build job, installed and started on a real PC, not yet through a whole meeting there.
• Neither installer is signed by a registered developer (see the install steps). On the Mac the keychain asks again after every update, because each build has a new signature.
• On Windows the API keys and mail password are kept in the settings file of your user profile, not in the Credential Manager yet; only WAV recordings can be imported; transcription on the computer runs on the CPU.
• Outlook.com accounts cannot send the mail: Microsoft no longer accepts passwords over SMTP for them.

SHA-256 of ZillaNote_0.3.1_aarch64.dmg: da30f336c671d6108a9ba2292aede0ab84261443342590b3d84ec6f573a1d7b8

SHA-256 of ZillaNote_0.3.1_x64-setup.exe: 0a71e4a30a9cb74092e404e9d187e14ae361ce5e3b7d6e609db0d97249a799c3 (built from this tag by the repository's build job, which also ran the unit tests on Windows, started the built app there and measured its bar)

The same two files are also attached as ZillaNote_aarch64.dmg and ZillaNote_x64-setup.exe, the names that the download buttons on the website and in the README always fetch from the latest release.

ZillaNote lite 0.3.0

Record a meeting, get minutes. ZillaNote lite records on your computer, finds who spoke, transcribes locally with Qwen3-ASR, writes the minutes with the language model you point it at, and mails them to you. The audio never leaves the computer; the transcript goes only to the language model endpoint you choose.

New in 0.3.0

• Open source. The code is under MIT or Apache-2.0, at your choice. THIRD-PARTY-NOTICES.md, shipped in every installer, lists what the app is built on, ships with and downloads, with their licenses. The two speaker models now come from this repository's own speaker-models-1 release, with the copies in the repository they were first taken from as fallback.
• It knows when a call starts and ends. Every call program (Teams, Zoom, Tencent Meeting, a browser on Meet) holds the microphone for exactly as long as the call, and the system says who holds it. When one opens the microphone, the bar's record button pulses and the window says "Meeting starts". When it lets go during a recording, the recording stops half a minute later as if you had pressed the button, after a countdown in the bar and in the window with Keep recording a click away; a program back within that half minute is the call going on. A program counts as a call only after a minute with the microphone, so dictation and voice messages change nothing. Nothing is installed and no permission is asked; both are switches in Settings. Nothing ever starts a recording by itself.
• Two speech models. Settings offers Qwen3-ASR 1.7B at 8 bits and 0.6B at 4 bits. A settings file naming one of the two choices in between from 0.2 is read as the nearest of these; a 1.7B 4-bit installation moves to the 8-bit model and downloads it once.
• Windows: the side bar at its real width (from 0.2.2), the same 38 by 142 points as on the Mac.

Install on a Mac

1. Apple silicon, macOS 12 or later (14.2 or later for the computer's sound and for noticing calls).
2. Open ZillaNote_0.3.0_aarch64.dmg and drag ZillaNote to Applications.
3. The app is not signed with an Apple Developer ID, so macOS blocks the first start of a downloaded copy: open System Settings → Privacy & Security, find "ZillaNote was blocked", press Open Anyway. Or in Terminal: xattr -dr com.apple.quarantine /Applications/ZillaNote.app
4. Allow the microphone, and "System Audio Recording" when the first recording asks.

Install on Windows

1. Windows 10 or 11, x64. Run ZillaNote_0.3.0_x64-setup.exe; it installs for the current user and needs no administrator.
2. The installer is not signed, so SmartScreen says "Windows protected your PC": choose More info, then Run anyway.
3. If Windows asks, allow the microphone for desktop apps (Settings → Privacy & security → Microphone).

Then, on either

The window offers to download the speech model and the speaker models (32 MB). In Settings, enter the language model for the minutes (any OpenAI-compatible endpoint: LM Studio, Ollama, or a hosted API with your key) and, if you want the minutes mailed, the mail account.

Known limits

• The call watch has been checked against a test program that opens the microphone, not yet through a whole Teams or Zoom call on either system; the first real calls will tell whether half a minute and a minute are the right numbers. Please report what you see.
• Windows is a preview: built, unit-tested and started on a Windows machine by the build job, installed and started on a real PC, not yet through a whole meeting there.
• Neither installer is signed by a registered developer (see the install steps). On the Mac the keychain asks again after every update, because each build has a new signature.
• On Windows the API key and mail password are kept in the settings file of your user profile, not in the Credential Manager yet; only WAV recordings can be imported; transcription runs on the CPU.
• Outlook.com accounts cannot send the mail: Microsoft no longer accepts passwords over SMTP for them.

SHA-256 of ZillaNote_0.3.0_aarch64.dmg: 685e182134d35de3b48f8363441c2569036bf7aac25d4b7a452436f2929992c5

SHA-256 of ZillaNote_0.3.0_x64-setup.exe: de0026ca78c55832957d5e0c2cc01d76b6284d8a5a506e9a9878d0df4cf4f44f (built from this tag by the repository's build job, which also ran the unit tests on Windows, started the built app there and measured its bar)

The same two files are also attached as ZillaNote_aarch64.dmg and ZillaNote_x64-setup.exe, the names that the download buttons on the website and in the README always fetch from the latest release.

Speaker models

The two ONNX files ZillaNote downloads to tell speakers apart, hosted here so that a change in the repository they were first taken from cannot break the download. The app checks each against its pinned SHA-256.

• segmentation.onnx (5,986,908 bytes): pyannote segmentation-3.0, MIT, © pyannote (Hervé Bredin). SHA-256 057ee564753071c0b09b5b611648b50ac188d50846bff5f01e9f7bbf1591ea25.
• embedding.onnx (26,543,975 bytes): the ONNX export of WeSpeaker's ResNet34 speaker embedding (wespeaker-voxceleb-resnet34-LM) as packaged by pyannote, CC BY 4.0. SHA-256 841645a9b5109369858ca89695e8b23ecfd4486b30bdd5cb3f99fef8d1c29205.

Both files are byte for byte the ones in fastrepl/anarlog at commit b9146a70, which stays the app's fallback source.

ZillaNote lite 0.2.2

Record a meeting, get minutes. ZillaNote lite records on your computer, finds who spoke, transcribes locally with Qwen3-ASR, writes the minutes with the language model you point it at, and mails them to you. The audio never leaves the computer; the transcript goes only to the language model endpoint you choose.

New in 0.2.2

• Windows: the side bar at its real width. On a PC the floating bar came out several times as wide as on the Mac, because Windows has a smallest width of its own for a window. The bar is now the same 38 by 142 points on both; the build job measures it on Windows and fails on any other size. Nothing else changed, and nothing changed on the Mac but the version number.

New in 0.2.1

• Windows (preview). An installer for Windows 10 and 11 on x64. It is built, unit-tested and started on a Windows machine by this repository's build job. On a real PC it has been installed and started; a whole meeting there (recording, transcription, minutes) has not been reported yet: expect rough edges and please report them. The far side of a call is recorded through WASAPI loopback, as on the Mac.
• A new look: graphite and violet. A violet ear on a graphite icon, a graphite side bar that reads over any window, a cool grey window with the same violet. The record button is a violet sphere; while recording it is a graphite button with a red stop square inside a violet halo that moves with the voice. The version is shown after the name.
• Mail that explains itself. "Send a test email" leaves the server's answer on the page and in the log. A refused password comes with what that provider wants instead (Gmail: an app password, with where to make one; QQ and 163: an authorization code). A Google app password pasted with its spaces works.
• A subject that says something. "Minutes 2026-09-20: Budget and revenue": the day without the time, then the meeting's name if you gave it one, else the first heading of the minutes that names a topic.
• Smallest model first. A new installation starts with Qwen3-ASR 0.6B at 4 bits (0.9 GB to download, 1.6 GB of memory while transcribing). The bigger models are one click away in Settings, each with its size, download time on your connection, and memory.

Install on a Mac

1. Apple silicon, macOS 12 or later (14.2 or later for the computer's sound).
2. Open ZillaNote_0.2.2_aarch64.dmg and drag ZillaNote to Applications.
3. The app is not signed with an Apple Developer ID, so macOS blocks the first start of a downloaded copy: open System Settings → Privacy & Security, find "ZillaNote was blocked", press Open Anyway. Or in Terminal: xattr -dr com.apple.quarantine /Applications/ZillaNote.app
4. Allow the microphone, and "System Audio Recording" when the first recording asks.

Install on Windows

1. Windows 10 or 11, x64. Run ZillaNote_0.2.2_x64-setup.exe; it installs for the current user and needs no administrator.
2. The installer is not signed, so SmartScreen says "Windows protected your PC": choose More info, then Run anyway.
3. If Windows asks, allow the microphone for desktop apps (Settings → Privacy & security → Microphone).

Then, on either

The window offers to download the speech model and the speaker models (32 MB). In Settings, enter the language model for the minutes (any OpenAI-compatible endpoint: LM Studio, Ollama, or a hosted API with your key) and, if you want the minutes mailed, the mail account.

Known limits

• Neither installer is signed by a registered developer (see the install steps). On the Mac the keychain asks again after every update, because each build has a new signature.
• On Windows the API key and mail password are kept in the settings file of your user profile, not in the Credential Manager yet; only WAV recordings can be imported; transcription runs on the CPU.
• Outlook.com accounts cannot send the mail: Microsoft no longer accepts passwords over SMTP for them.

SHA-256 of ZillaNote_0.2.2_aarch64.dmg: 015db37005b1aae12a9774999f9fce9c1229db68ee365d131677273353fb91fb

SHA-256 of ZillaNote_0.2.2_x64-setup.exe: 2804ac22f8ed5dedb699ace438b1b88135240c769a7b8e5d082d6fc745b700d2 (built from this tag by the repository's build job, which also ran the unit tests on Windows, started the built app there and measured its bar)

ZillaNote lite 0.2.1

Record a meeting, get minutes. ZillaNote lite records on your computer, finds who spoke, transcribes locally with Qwen3-ASR, writes the minutes with the language model you point it at, and mails them to you. The audio never leaves the computer; the transcript goes only to the language model endpoint you choose.

New in 0.2.1

• Windows (preview). An installer for Windows 10 and 11 on x64. It is built, unit-tested and started on a Windows machine by this repository's build job, and has not yet been used for a real meeting on a PC: expect rough edges and please report them. The far side of a call is recorded through WASAPI loopback, as on the Mac.
• A new look: graphite and violet. A violet ear on a graphite icon, a graphite side bar that reads over any window, a cool grey window with the same violet. The record button is a violet sphere; while recording it is a graphite button with a red stop square inside a violet halo that moves with the voice. The version is shown after the name.
• Mail that explains itself. "Send a test email" leaves the server's answer on the page and in the log. A refused password comes with what that provider wants instead (Gmail: an app password, with where to make one; QQ and 163: an authorization code). A Google app password pasted with its spaces works.
• A subject that says something. "Minutes 2026-09-20: Budget and revenue": the day without the time, then the meeting's name if you gave it one, else the first heading of the minutes that names a topic.
• Smallest model first. A new installation starts with Qwen3-ASR 0.6B at 4 bits (0.9 GB to download, 1.6 GB of memory while transcribing). The bigger models are one click away in Settings, each with its size, download time on your connection, and memory.

Install on a Mac

1. Apple silicon, macOS 12 or later (14.2 or later for the computer's sound).
2. Open ZillaNote_0.2.1_aarch64.dmg and drag ZillaNote to Applications.
3. The app is not signed with an Apple Developer ID, so macOS blocks the first start of a downloaded copy: open System Settings → Privacy & Security, find "ZillaNote was blocked", press Open Anyway. Or in Terminal: xattr -dr com.apple.quarantine /Applications/ZillaNote.app
4. Allow the microphone, and "System Audio Recording" when the first recording asks.

Install on Windows

1. Windows 10 or 11, x64. Run ZillaNote_0.2.1_x64-setup.exe; it installs for the current user and needs no administrator.
2. The installer is not signed, so SmartScreen says "Windows protected your PC": choose More info, then Run anyway.
3. If Windows asks, allow the microphone for desktop apps (Settings → Privacy & security → Microphone).

Then, on either

The window offers to download the speech model and the speaker models (32 MB). In Settings, enter the language model for the minutes (any OpenAI-compatible endpoint: LM Studio, Ollama, or a hosted API with your key) and, if you want the minutes mailed, the mail account.

Known limits

• Neither installer is signed by a registered developer (see the install steps). On the Mac the keychain asks again after every update, because each build has a new signature.
• On Windows the API key and mail password are kept in the settings file of your user profile, not in the Credential Manager yet; only WAV recordings can be imported; transcription runs on the CPU.
• Outlook.com accounts cannot send the mail: Microsoft no longer accepts passwords over SMTP for them.

SHA-256 of ZillaNote_0.2.1_aarch64.dmg: 0df881bd732d95444ec43310f445ec459d6b778467a3ad9cb671719f92afece3

SHA-256 of ZillaNote_0.2.1_x64-setup.exe: bbc70aa169925d2955b37d8e4564a81d554d9317d581c1f52768724453f3ae5c (built from this tag by the repository's build job, which also ran the unit tests on Windows and started the built app there)

ZillaNote lite 0.2.0

Record a meeting, get minutes. ZillaNote lite records on your Mac, finds who spoke, transcribes locally with Qwen3-ASR, writes the minutes with the language model you point it at, and mails them to you. The audio never leaves the Mac; the transcript goes only to the language model endpoint you choose.

New in 0.2

• Both sides of a call. The computer's own sound is recorded next to the microphone (macOS 14.2 or later), so the far side of a Teams or Zoom call is in the transcript. On loudspeakers, the echo of the far side is left out of what is transcribed.
• Speakers. Transcripts say who spoke. Name a speaker once and that voice is recognized in later meetings. Only naming teaches the app a voice, so a wrong guess cannot feed on itself.
• Speech models, downloaded in the app. Four to choose from (table below). Settings shows each one's download size, the time it takes on your connection, and the memory it needs. Downloads are pinned to a revision, checked by SHA-256, continue where they stopped, and have a mirror for every file. Files LM Studio already has are reused.
• Better minutes. A new default prompt with the shape spelled out: numbered topics, decisions, next steps with the owner in brackets, and what was raised and left open. Minutes are written by themselves once the transcript is ready (a switch in Settings), tried again after a passing failure, and caught up at the next start if the app was closed in between.
• Everyday things. A menu-bar icon, import of recordings made elsewhere (button or drop), the API key and mail password in the macOS keychain, a smaller floating bar, a proper way to quit that always closes the recording into a whole file, and a log file for the day something did not happen.
• A new icon and colors.

Speech models

Measured on an M4 Pro, against the 1.7B 8-bit transcript of the same audio.

Model	Download	Memory while transcribing	3 minutes of audio	Same words, clear English	Same words, mixed English and Mandarin
1.7B 8-bit (default)	2.8 GB	3.5 GB	10 s	reference	reference
1.7B 4-bit	1.9 GB	2.6 GB	8 s	98.8%	97.9%
0.6B 8-bit	1.2 GB	1.85 GB	5 s	97.7%	93.0%
0.6B 4-bit	0.9 GB	1.55 GB	4 s	97.9%	91.6%

The memory is used only while a recording is being transcribed; the app idles at about 100 MB.

Install

1. You need a Mac with Apple silicon, macOS 12 or later (14.2 or later for the computer's sound).
2. Open ZillaNote_0.2.0_aarch64.dmg and drag ZillaNote to Applications.
3. The app is not signed with an Apple Developer ID yet, so macOS blocks the first start of a downloaded copy. Open System Settings → Privacy & Security, find "ZillaNote was blocked", and press Open Anyway. Or, in Terminal: xattr -dr com.apple.quarantine /Applications/ZillaNote.app
4. Allow the microphone, and "System Audio Recording" when the first recording asks for it.
5. The window offers to download the speech model you chose and the speaker models (32 MB).
6. In Settings, enter the language model for the minutes (any OpenAI-compatible endpoint: LM Studio, Ollama, or a hosted API with your key) and, if you want the minutes mailed, the mail account.

Known limits

• Not notarized. Besides the first-start step above, the keychain asks for permission again after every update, because each build has a new signature.
• macOS on Apple silicon only.

SHA-256 of the installer: c5fdee36f64eff498e1057b9e6ec2f4968261da99e84193aadf9a0f085a49769