Choosing Whisper: include the time you spend editing

Compare two models on your own material

Sonano offers Tiny, Base, Small, Medium, Large-v3 Turbo and Large-v3 variants of Whisper. A smaller model can be a starting point; compare a larger variant when recordings are difficult. Bring your actual language and source material to the test.

Whisper’s published speed table measures English transcription on an A100. It illustrates tradeoffs, but it cannot tell you how long your Mac will take to process an interview. Model size, runtime and the recording itself affect the result.

Count correction time too

Choose a representative excerpt with names, numbers, an accent and some background sound. Run the same file through two candidate variants. Note both the waiting time and the corrections needed. A quick result that needs rewriting sentence by sentence may take longer to finish.

• Text: are there omissions, repetitions or incorrect names and numbers?
• Timing: do subtitle changes land near the matching speech?
• Editing: how much punctuation, segmentation and terminology needs work?

Upstream describes Turbo as an optimization of Large-v3 with a speed and accuracy tradeoff. Performance still varies by language. Include it as a candidate; no single variant is the best fit for every recording.

Review against the audio

Import audio or video into Sonano’s transcription workflow and select the model and language. When the result is ready, replay it in sections. Check names, amounts, dates and specialist terms first. Around silence, music or unclear speech, look for text that was never spoken.

Recognition, translation and identifying speakers are separate tasks. The upstream Whisper project’s capabilities do not mean every option is exposed for every variant in Sonano. This guide covers transcription and segment timing in the app.

Check subtitles in their destination

For a plain transcript, review the text and export TXT. For subtitles, check wording, timing and line length before exporting SRT or VTT, then replay them in the target player. Long sentences can be split for reading; very brief captions may disappear too quickly.

Exporting a subtitle file does not burn captions into your video. If narration or an audiobook is your next step, use the corrected transcript as source text. Review first so recognition mistakes do not become spoken mistakes in the next output.

See the Whisper model overview and transcription documentation for the current app options.