Mockingbird Studio

A local transcription workstation for Windows x64, using Whisper and Canary to recognize speech independently and help you review uncertain wording.

Source: github.com/PufferfishGaming/Mockingbird

Version 0.1.22 — release candidate

New in 0.1.22: search every transcript at once, meeting summaries, the web page as an app on a phone with a QR code to open it, a Server window with Studio's own setup pages, Resume and a watch folder in the Client, a button that opens the recordings folder, and a livelier look.

Editions

Server and Client are new in 0.1.20. Together they do what Studio does: the Server holds and tunes the models, the Client (or the web page) does the rest. Each edition installs, updates and keeps its data on its own (TriASR, TriASR-Server, TriASR-Client), and has its own checksum file (SHA256SUMS.txt, SHA256SUMS-server.txt, SHA256SUMS-client.txt) and update file (latest.json, latest-server.json, latest-client.json).

Quick install

Open PowerShell (Windows 10 or 11, x64) and paste one line. It downloads the installer from this project's download release, checks it against the published SHA256 file, and opens the setup wizard. No administrator permission is needed.

& ([scriptblock]::Create((irm https://raw.githubusercontent.com/PufferfishGaming/Mockingbird/main/scripts/install.ps1))) -Edition Studio

Use -Edition Server or -Edition Client for the other two. Add -Quiet to install without any window (choosing it means you accept the license, GPL-3.0-or-later), -NoLaunch to not start the program afterwards, or -DownloadOnly to stop after the check. The command runs scripts/install.ps1 from this repository, so read it first if you like; it changes nothing else on the computer. The installer is not signed and a file downloaded this way is not marked as coming from the internet, so Windows SmartScreen does not warn about it: the SHA256 check is what the script gives you instead. If the checksum does not match, the file is deleted and nothing is installed.

Install

Download Mockingbird-Studio-Setup.exe from the download release and double-click it. The setup wizard installs for your user only and includes the .NET runtime, so nothing else is required. Download models inside the app after installing.

The Server edition downloads its models after installing, like Studio; the Client needs none.

To remove the app, use Settings → Apps → Installed apps → Mockingbird Studio. Uninstalling keeps your projects, settings and models.

Installing an unsigned program

The installers are not code-signed yet. Windows trusts a program it has not seen often only when it carries a code-signing certificate, which costs money every year, so for now Windows warns before Mockingbird is installed. The warnings do not mean anything was found in the file: Mockingbird is open source, and every release publishes the SHA256 of each installer, so you can check that your download is exactly the file built here.

1. Check the download (recommended). Open PowerShell in the folder you saved it to and run Get-FileHash .\Mockingbird-Studio-Setup.exe. The hash must equal the line for that file in SHA256SUMS.txt from the same release (SHA256SUMS-server.txt and SHA256SUMS-client.txt for the other editions). If it differs, delete the file and download it again.
2. If the browser holds the download back. Edge may say the file "isn't commonly downloaded": point at the download, click … → Keep, then Show more → Keep anyway. Other browsers ask in a similar way; choose to keep the file.
3. If "Windows protected your PC" appears when you open the installer (Microsoft Defender SmartScreen): click More info, then Run anyway. The setup wizard opens.
4. If "Smart App Control blocked an app" appears (Windows 11 with Smart App Control on): there is no Run anyway button, and Smart App Control cannot make an exception for one program; it only lets through signed programs or ones Microsoft already knows. Mockingbird can be installed only with Smart App Control turned off (Windows Security → App & browser control → Smart App Control settings → Off). That is your choice to make, since it switches the protection off for every program; on Windows versions from before the 2026 change it cannot be switched on again without reinstalling Windows. Otherwise, wait for a signed release.

Instead of steps 2 and 3 you can install with the PowerShell line under Quick install: it checks the SHA256 for you, and a file it downloads is not marked as coming from the internet, so SmartScreen does not ask. You can also clear that mark yourself after checking the hash: right-click the installer → Properties → tick Unblock → OK (or Unblock-File .\Mockingbird-Studio-Setup.exe). Updates made from inside the app check the SHA256 the same way and do not show these warnings.

Updates

From version 0.1.16 the app checks GitHub for a newer release when it starts (at most every 12 hours) and shows a banner with Update now, Later and Skip this version. Updating downloads the new Mockingbird-Studio-Setup.exe, checks its SHA256 against the published latest.json, closes the app, installs it and reopens it. It never updates in the middle of a transcription, model download or benchmark, and your projects, settings and models are kept. Turn the automatic check off, or check manually, in Settings → Updates. Each update downloads the full installer; there are no partial updates yet. Earlier versions must be updated once by hand.

Native Linux support is pending; there is currently no native Linux build.

Features

• Audio/video import with progress, cancellation and resumable jobs: a recording that was stopped or failed is resumed where it stopped (Resume in Studio, the Server window, the Client and the web page; the stages already finished are reused).
• A watch folder: recordings added to a chosen folder are transcribed one after another and the transcript is saved next to each one, as text or subtitles. Studio transcribes them itself; the Client sends them to the connected server (they wait while there is no connection).
• Whisper's 100-language catalog and independent Canary comparison for 25 languages.
• Waveform playback and manual transcript review, with the places where the two engines disagree marked for listening. A local correction model chooses between them where it is at least 90% sure (on by default; it can be switched off in Settings).
• Removal of Whisper's runaway repetition, and an optional "Skip silence and music" setting for recordings with long quiet or noisy stretches (it removes songs, so it is off by default).
• TXT, Markdown, JSON, CSV, SRT, VTT and DOCX export.
• Persistent model downloads, CPU/Vulkan backends, optional compatible CUDA/ROCm runtimes and measured auto-tuning.
• Live activity output, an interactive PowerShell panel and light/dark/system themes. Buttons, list rows and choices light up under the pointer and press in when clicked, the field or button the keyboard is on is clearly marked, and pages fade in (only when Windows' animation effects are on).
• Karaoke-style playback in the review: while the recording plays, the words already said are coloured, the word being said is bold, and the selection follows the recording (in Studio, the Client and the web page).
• A history of your transcripts: the Projects page (in Studio, the Client and the web page) lists every project with its date and state; click a finished one to open its transcript in Review, and delete a project you no longer need (its transcript, edits and working files go; the recording file you chose stays).
• Search every transcript at once: type in the box on the Projects page (Studio, the Client and the web page) and every project whose transcript or name holds the words is listed with its matching passages, the words picked out; a click opens the transcript right there. Capital letters, accents and extra spaces do not matter ("kotelezo" finds "kötelező").
• Summaries: open a transcript, unfold Summary and press Summarize (Studio, the Client and the web page). A language model on this computer, or on the server, writes what the conversation was about, the key points, the decisions, who does what and the open questions, in the language of the transcript and with the names given to the speakers; copy it or save it as Markdown. It is a draft to check against the transcript. Studio uses its correction model (Qwen3 4B) for it; for better summaries, above all in Hungarian, download the optional Summary model (Gemma 3 12B, 7.3 GB, Gemma Terms of Use) on the Models page. Measured on English meetings and a Hungarian podcast before it was built: with Gemma every checked statement held, with the small model one to three per summary were wrong. A summary takes seconds on a graphics card and about a minute on the processor.
• An in-app recorder: record with the microphone in Studio, in the Client or in the browser; the recording becomes the file to transcribe or to send. Open folder on the recorder opens the folder the recordings are kept in.
• Dictation: a small window floats above your other programs (it never takes the keyboard from them); start it with its button or with keys (Ctrl+Alt+Space to begin with; press Change keys, press the keys you want, click Done), speak, and every phrase is typed (or pasted) where your cursor is, in any program. Studio reads the phrases on this computer; the Client, and Studio's Remote server page, send each phrase to the server. It needs a Whisper model; it types only into programs that are not running as administrator.
• Two languages: if you speak (or a recording switches between) two languages, English and Hungarian say, choose both, and each part is written in the language it was spoken in, in dictation, notes and recordings alike. With one language chosen, whatever is said in another one is written translated into it, and auto-detection can mistake an accent for another language. A phrase the program is not sure about is read in both languages and takes a moment longer; a recording is divided into stretches of one language, and each is transcribed in its own. A recording that turns out to be in one of the two only is transcribed exactly as with that language chosen.
• Speakers: choose Tell speakers apart (or how many people speak) when you start a transcription, and every part of the transcript is marked with its speaker, cut where the speaker changes; the exports say who speaks ("Speaker 1: ...", WebVTT voice tags, a CSV column). Studio's watched folder has the same choice for the recordings it picks up. In the review (Studio, the Client and the web page) give the speakers their names, Anna instead of Speaker 1: they show in the transcript at once, are saved with the edits and are used in every export. It runs on the processor and takes a few percent of the recording's length. Measured on English meetings of four people, 95 % of the words went to the right speaker.
• Notes with live transcription: a Notes page in Studio (this computer's notes) and a Notes tab on a server's pages (the Client, Studio's Remote server page and the web page; the notes are kept on the server, so every window sees the same ones). A note saves itself as you type; press Record and every phrase you say is added to the end of it. Keys of your own choosing start and stop recording from any program in Studio and the Client (and while the page is open in the browser), and a little window shows while they record. A note that was changed on another computer is never overwritten: what you wrote is kept as a note of its own.
• Transcribe a link: paste the address of a video, a podcast episode or an audio file on any site (not only one video site) in Studio, in the Client or in the browser. A link straight to an audio or video file just works; for web pages Mockingbird can install a helper program (yt-dlp, downloaded once from its project and checked against its published checksum). A server fetches links for the computers that use it only when it has a password.
• The interface in English, Hungarian, German, Spanish and French: chosen on the first start, switchable in Settings without a restart.
• In-app update checks with a verified one-click update.
• Three editions: Studio does everything on one computer, Server holds the models and serves them, Client is only the window (see Editions).
• A server you can name and protect with an optional password, found on your network by the computers that want to use it, with every connection over the network encrypted (see Servers on the network).
• An optional HTTP API, so that other programs can send recordings and fetch transcripts.

Recognition can be wrong, particularly with music, noise or silence. Review important transcripts. Readable export normalizes spacing without rewriting wording.

Servers on the network

In Studio the Servers page (in the sidebar, below Models) lists the servers found on your network, and next to it you can host one of your own. The Server edition opens on that hosting box, with the recordings it was sent (cancel, resume or delete them there), and has Studio's Models, Settings and Advanced pages (languages, backends, tuning, diagnostics, terminal) to set the server up; the Client edition has only the list, in a panel on the right.

Hosting. Press Start hosting (it becomes Stop hosting while the server runs). Give the server a name (shown to the others; empty uses the computer's name) and, if you like, a password: type one or press Make a password; with none, anyone who can reach the server may use it. Who can use it chooses between This computer only and Computers on the network; with the first only programs on the same computer can connect. The server runs while the app is open, and starts again next time if you left it on. A newly installed Server edition does not host until you press Start hosting. The first time on the network, Windows may ask whether to let it through its firewall.

Encrypted. Every connection over the network is encrypted (TLS). The server makes a certificate for itself, and its fingerprint is shown in its Identity box. The first time a computer connects it shows the same fingerprint; compare the two, and confirm only if they match. After that the server must show the same fingerprint, or the connection is refused with a warning. The password is sent only after you have confirmed the fingerprint, and it can be remembered (protected by your Windows account) or asked every time. New identity makes a new certificate; every computer then asks again.

Finding servers. A server reachable from the network announces its name and address on the local network every two seconds; the list shows what it hears. A server on another network is added by typing its address (192.168.1.20 or kitchen:8642). The announcement is only a hint and is never trusted without the fingerprint check.

Using a server. Choose it and press Connect. In Studio a Remote server page appears in the sidebar, in the Client it fills the window: New sends a recording and a language to the server, Projects follows the recordings and their progress, and Review opens a transcript with its audio, the two engines' wording and your edits, which are saved on the server and can be exported. The speech programs run on the server; the client only sends and reads.

Mockingbird Client Webview

Every hosted server also serves a web page, Mockingbird Client Webview, at its own address, for computers that do not have Mockingbird installed. Open the address in a browser (Open web page in the hosting box does it for this computer), enter the password if there is one, and you can send a recording, follow the recordings on the server, review and edit a transcript with its audio, and export it: the same pages as the Client, in the five languages of the program. On the network the address starts with https://. The certificate is the server's own, so the browser warns the first time; compare the fingerprint in the certificate details with the one in the server's Identity box before you continue. The page loads nothing from any other address, sets no cookies and keeps the password only for the browser tab. A server without a password only answers requests addressed to its IP address, localhost or its computer's name.

On a phone the page works like an app: the tabs are a bar at the bottom with icons, the fields are large enough to tap, and nothing scrolls sideways. Point the phone's camera at the QR code in the hosting box (Studio's Servers page or the Server window; it appears while the server is on with Computers on the network), or at the one behind Open on a phone in the Client and on the web page opened on a computer, with the phone on the same Wi-Fi; or type the server's https:// network address. Accept the certificate once as above, then put it on the home screen: on an iPhone Share → Add to Home Screen, on Android ⋮ → Add to Home screen (or Install). It then opens full screen with its own icon. Recording with the phone's microphone needs the https:// address; on a plain http:// address the page says so and links to the encrypted one.

The HTTP API

The same server answers programs. On this computer http://127.0.0.1:8642 works without encryption; from the network use https:// (a self-signed certificate, so curl.exe -k, or pin the fingerprint). When a password is set, every request except the health check needs it (Authorization: Bearer <password> or X-Api-Key). Recordings sent through the API are kept in the projects folder (Api/Incoming) and listed under Projects; the API shows only what was sent through it, and works on one recording at a time.

curl.exe http://127.0.0.1:8642/v1/health
curl.exe -X POST --data-binary "@meeting.mp3" -H "Authorization: Bearer PASSWORD" "http://127.0.0.1:8642/v1/transcriptions?language=auto&name=meeting.mp3"
curl.exe -H "Authorization: Bearer PASSWORD" "http://127.0.0.1:8642/v1/transcriptions/ID?wait=60"
curl.exe -H "Authorization: Bearer PASSWORD" "http://127.0.0.1:8642/v1/transcriptions/ID/transcript?format=srt"

Errors are {"error":{"code":"...","message":"...","type":"..."}}.

Privacy

Transcription runs locally. Requested model/runtime downloads and the update check (which can be turned off) contact their providers. A hosted server listens only when you switch it on, encrypts what it sends over the network and announces itself on your local network only while the network option is on. The microphone is used only while recording, dictation or a note is being recorded; dictated words are typed into whichever program has the keyboard. Terminal commands can access the network. Projects and logs can contain private information. See PRIVACY.md, also available in Settings.

Build

Install the .NET SDK specified in global.json. From PowerShell:

./scripts/build.ps1
./scripts/test.ps1

To look at what you built, ./scripts/try-edition.ps1 starts all three editions (or one, with -Edition Studio, Server or Client); Studio uses your usual data, Server and Client keep theirs under artifacts/try.

For complete app smoke checks, supply the Windows native runtimes under Runtimes/ and run scripts/verify.ps1. Packaging additionally requires the app-local Visual C++ runtime inputs:

./scripts/package.ps1
./scripts/build-msi.ps1
./scripts/build-setup.ps1

build-setup.ps1 wraps the MSI in the setup wizard and needs the WiX bootstrapper extension (WixToolset.BootstrapperApplications.wixext 6.0.2) placed under .tools/wix-extensions; the script prints the exact path if it is missing.

Version metadata comes from Directory.Build.props. Preserve both normal and publish package lock files. Native runtime binaries and model weights are external prerequisites and are excluded from Git.

Legacy TriAsr.* names, storage paths and TRIASR_* environment variables remain for upgrade compatibility. Upgrades and uninstall preserve user projects, settings and models.

License and release status

Copyright © 2026 PufferfishGaming. Original source is GPL-3.0-or-later; see LICENSE. Dependencies retain their own licenses; see THIRD_PARTY_NOTICES.md.

This candidate is unsigned. FFmpeg, the audio converter inside Studio and Server, is Mockingbird's own LGPL build of the official FFmpeg source, and its complete source (FFmpeg-<version>-source.zip) is published on the release page with the installers (scripts/build-ffmpeg.ps1 builds it). Before public binary distribution, complete the review of the other third-party notices, clean-machine installation tests and remaining accuracy/accessibility acceptance checks.