Turn a recording into editable subtitles locally with Buzz

Buzz provides a desktop route from an audio or video file to editable subtitles using a local speech-recognition model. Start with a short recording that you have permission to process. The intended result is an SRT or VTT file with corrected words and usable timings, not merely a block of transcript text.[1][4]

The application supports several engines, including local Whisper implementations and an OpenAI API option. Choose a local engine for the workflow described here; selecting a cloud API changes where processing takes place. A local model may need an initial download before it can run on the computer.[1]

The steps below follow Buzz’s documentation. The timing example is an original calculation; no transcription-speed or word-accuracy result is claimed here.

Import the recording and choose the job

1. Install Buzz using the download route linked from its official installation instructions. This walkthrough is intended for an Apple-silicon Mac or Windows computer, not the iPhone itself.
2. Open File → Import Media File, or use the import button. Select the audio or video you want to caption.
3. Choose a local engine such as Whisper. For a short first exercise, our suggested starting model is small; select Transcribe to retain the spoken language. This is a starting choice, not a benchmarked hardware recommendation.
4. Set the spoken language when it is known, then run the job. When it completes, open the finished transcription to inspect the timed segments.

Whisper’s original release includes multilingual speech recognition and translation into English. That distinction is worth choosing deliberately: captions of the original speech and an English translation are different deliverables. Neither should be labelled as the other just because both are readable text.[5]

Correct words while listening, not from memory

Buzz’s transcription viewer lets you edit segment text and timestamps while playing the associated audio. It also supports slower playback and repeated listening to a segment. Work through proper names, numbers and any section where speech overlaps, using the recording as the reference.[4]

For a first pass, our suggested order is meaning before typography. Correct an omitted “not” or an incorrect amount before polishing punctuation. Then divide a long thought into readable phrases. A beautiful line break cannot repair a sentence that says the opposite of the speaker.

A short cue can contain too much text

Take the invented subtitle “Please bring the blue folder.” It contains 29 characters, counting spaces and punctuation. Displayed for one second, its reading rate is 29 characters per second; displayed for two seconds, it is 14.5. The wording has not changed, but the reading task has.

For a simple timing screen, divide each cue’s character count by its duration. A chosen threshold such as 20 characters per second can flag cues for review; it is an illustrative setting, not a universal accessibility standard. Also inspect overlaps, malformed timestamps and durations of zero or less. These checks can find timing problems, but listening is still needed to judge the words.

Timing edits must respect the recording. If the next person begins speaking, simply extending the first cue may create an overlap or attach the wrong words to the wrong moment. Split or reshape the caption at a sensible phrase boundary, then replay the transition. This is a suggested editing method, not a claim about a particular recorded interview.

Export the file the next application needs

Use the viewer’s export controls to save SRT or VTT for timed captions. TXT is useful for the words alone but is not a substitute for the timing information. Keep the original media and the corrected subtitle file together so that the result can be reopened and checked later.[4][3]

Finally, load the subtitles beside the original recording in the intended player or editor. Check the beginning, a middle transition and the final cue for synchronisation. A file that passes a timestamp check can still be offset against the video. The completed job is a caption track that says the right thing at the right moment—not just a successful model run.

Report an error or suggest a correction