Docs / Meeting Recording / Whisper Models

Whisper Models

Choose between speed and accuracy for local transcription.

QuickContract uses whisper-cpp, a high-performance C++ implementation of OpenAI's Whisper speech recognition model, to transcribe meeting audio locally on your Mac. The whisper-cpp binary runs as a sidecar process alongside QuickContract — it launches automatically when a recording starts and shuts down when the recording ends.

You need to download at least one Whisper model before you can record. QuickContract offers two model sizes, each with different tradeoffs between transcription speed and accuracy.

Model comparison

Model Size Speed Accuracy Best for
Base ~140 MB Fast — real-time on most Macs Good Day-to-day meetings, English conversations, quick recordings
Large ~3 GB Slower — may lag on older hardware Highest Multilingual calls, heavy accents, critical transcriptions where every word matters

The base model processes 10-second audio chunks faster than real time on Apple Silicon Macs, which means transcription keeps pace with the conversation. The large model is significantly more accurate, especially for non-English languages and technical vocabulary, but requires more processing time and memory.

Start with the base model

It's fast enough for real-time transcription and accurate for most English conversations. You can always download the large model later if you need higher accuracy for multilingual calls or specialized terminology.

Downloading a model

Open Settings

Click the gear icon in the sidebar or press Cmd + , to open Settings.

Go to Meeting Recording

Select the Meeting Recording section in the Settings panel.

Download your preferred model

Under Whisper Model, you'll see the available models with download buttons. Click Download next to the model you want. The download progress is shown inline. Once complete, the model is stored locally and ready to use.

Select the active model

If you have both models downloaded, select which one to use for future recordings. You can switch between them at any time — the change takes effect on the next recording.

How transcription works

When a recording starts, QuickContract launches the whisper-cpp sidecar binary with the selected model loaded into memory. Audio from both the system and microphone streams is fed to whisper-cpp in 10-second chunks. Each chunk is transcribed independently and appended to the running transcript with its speaker label ("me" or "them") and timestamp.

Because each chunk is small and self-contained, a transcription error in one segment doesn't affect subsequent ones. If a chunk contains silence or unintelligible audio, whisper-cpp simply returns an empty result and moves on.

Storage and management

Downloaded models are stored in QuickContract's application support directory. You can delete a model from the Settings panel to free up disk space. If you delete the active model, QuickContract will prompt you to download one before your next recording.