On this page
1. Prepare the recording and install Buzz #
Download Buzz through the official release links, choosing the package for your computer. The GitHub release page links the maintained download locations. On Windows, follow the package instructions: releases distributed as an EXE plus two BIN files need all three together.
Start with a recording you can play locally. Listen to its opening and check which language the speaker uses. A voice memo with a clear beginning makes a better first job than a video with several separate audio streams.
When Buzz opens, you should see its transcription list. You don't need to add a remote API key for this local workflow. For more background on its engine, Whisper.cpp is the local C/C++ project used by one of Buzz's backend choices.
2. Download a local recognition model #
Open Help → Preferences → Models. Download the local model you intend to use before working without a connection. The model preferences documentation explains how Buzz manages downloads and unused models.
For a starting job, I'd choose a modest model size that your computer can load. You can move to a larger model later if the text needs improvement. A larger model is a separate download, with its own memory requirements.
The offline setup FAQ also explains copying a prepared model cache to an offline computer. Finish that preparation while connected. The success check here is that the model is downloaded and available for selection, rather than waiting for it to arrive during an offline job.
3. Import the recording #
Choose File → Import Media File, use the toolbar +, or press Ctrl+O on Windows/Linux or Command+O on Mac. Select your saved audio file.
The import dialog should offer task, language and model settings. These are the controls in the official file-import instructions. You haven't started recognition yet, so this is the point to set the output you want.
Use a descriptive source filename, such as product-demo-voiceover.wav, before importing. That makes it easier to recognize the job in the list and connect its eventual transcript to the right recording. Avoid treating several similarly named recordings as interchangeable when you review corrections.
4. Set the language and transcription task #
Set Task to Transcribe to keep the text in the spoken language. Choose the language explicitly when you know it. Detect Language makes the recognizer infer it from the opening audio, which can be a poor starting point when the file begins with silence or music.
Select the local Whisper.cpp backend and the model you've prepared. An English-only model ending in .en is for English recordings; choose a multilingual model for another language. The task and language reference defines Transcribe and Translate to English.
For example, a Portuguese interview you plan to quote in Portuguese needs Transcribe with Portuguese selected. Translation is a different deliverable. Confirm the task before running so you don't spend processing time creating the wrong output.
5. Choose text or subtitles and add spelling hints #
Under Export As, select TXT for ordinary prose, SRT for a SubRip caption file or VTT for WebVTT. TXT is the documented default. Decide which format the next application needs before starting.
Enable Word-Level Timings if you want word-level segments for later subtitle regrouping. Leave it off for a straightforward text transcript. Extract speech is an optional preprocessing choice for separating speech from the recording; it isn't required for the basic import.
Open Advanced… and use Initial prompt for terms that are easy to misspell. A short list of product names or technical terms from the recording is more helpful than asking the recognizer to rewrite the speaker. For a website demo, that might include the application name and the names of features the narrator actually says.
You should now have the task, language, local model and chosen export format set together. Check them once before processing.
6. Run recognition and open the result #
Click Run. The file enters Buzz's job list. Wait until its status reads Completed, then double-click the row to open the transcription.
Read the opening sentences against the audio. Check a sentence near the middle and the end as well, so a recognizable opening doesn't hide a problem later in the file. Pay particular attention to names, quantities and instructions someone could act on.
The viewer documentation describes editing the transcript and using Resize to regroup segments. For subtitles, word-level timings give the regrouping tool more information. For prose, focus first on the words and speaker meaning.
7. Inspect the exported file before handing it over #
Open the TXT, SRT or VTT file created by the export choice. Check the actual saved output as well as the viewer. Buzz's preferences let you customize the default export filename using the input name, task, language, model and export time.
For TXT, confirm the text opens correctly in the document or notes application you plan to use. For captions, check that the timestamps and text appear with the matching recording in your video workflow. Make corrections before the transcript becomes an article quote or a public caption track.
If your next step is writing from the transcript, Whisper explains the underlying recognition project's setup and model choices. If the handoff calls for more document formats, Vibe is another free desktop option.
Known problems and fixes #
Speaker identification stops after a model download
A January 2026 report described speaker identification failing in Buzz 1.4.3. The maintainer traced the likely problem to the downloaded alignment models and suggested clearing the damaged model folder before retrying (issue 1367).
Open Help → Preferences → Models, select a downloaded model and use Show file location. Follow the issue's exact cache-folder guidance for that speaker-identification failure, restart Buzz and allow the download to finish. The maintainer later said a development fix was available. Ordinary file transcription doesn't require running speaker identification first.
A video has several tracks, but only one is transcribed
In February 2026, a user reported that importing a multitrack MP4 or MKV selected the first audio stream (issue 1386). The report's workaround was to extract the intended stream with FFmpeg and import that separate audio file.
Use that approach when a video has separate tracks for speakers or languages. Verify the extracted recording by listening before importing it. A model setting can't recover a speaker who isn't present in the audio stream being processed.
Troubleshooting table #
| Problem | Likely cause | Next action |
|---|---|---|
| The first offline job cannot start | Required model hasn't finished downloading | Complete the model download while connected, then retry locally |
| Recognition takes too long | Model too demanding for available hardware | Try a smaller local model, as Buzz's FAQ recommends |
| Text is in English instead of the spoken language | Translation task selected | Choose Transcribe and the spoken language |
| A product name is repeatedly misspelled | Recognition needs terminology context | Add the spoken term to Initial prompt and review the result |
| One speaker is missing from a multitrack video | Wrong source audio stream | Extract and listen to the intended track before import |
For a crash after an incomplete model download, the official FAQ recommends deleting and downloading the affected model again. Help → About Buzz → Show logs opens the log location for a detailed error report.
Variations #
Create captions with readable line groups
Select SRT or VTT under Export As and enable Word-Level Timings before recognition. After completion, open the viewer and Resize the segments. The documented options include maximum subtitle length and punctuation-based splitting when word timings are available. Read the regrouped lines alongside the video before using them.
Translate the recording into English
Change Task to Translate to English when English output is the goal. Choose a model that supports that task. Translation changes the words you can quote, so keep translated notes clearly identified in your writing workflow. A transcript in the original language is the better starting point for checking an exact quotation.
Repeat the job with a different model
If the transcript misses important words, prepare another model and import the same source again with that model selected. Compare the troublesome passages against the recording. Keep the exported filenames descriptive enough to tell the attempts apart; a file labeled only “final” won't explain which model produced it.
How to transcribe an audio file with Buzz: final checks #
- The intended local model finished downloading.
- The correct recording and spoken language were selected.
- The task produced the original language or intended English translation.
- Buzz showed Completed, and the transcript opened.
- Names, numbers and important statements were checked against the recording.
- The saved TXT, SRT or VTT opened in its destination application.
Next steps #
Use Buzz vs Vibe to choose between timed-segment editing and broader document handoffs. For interview editors, Linux notes and subtitle-focused options, see the best free offline transcription tools.
Featured illustration: WebIndexly, created with OpenAI image generation.



