Model selection
Choose the right Whisper model for the browser
Tiny minimizes download and memory use, Base offers a middle ground, and Small can improve difficult transcription at a much higher resource cost. English-only variants are available for English speech.
Quick answer
Whisper Online downloads a whisper.cpp model, processes the selected media locally, lets you correct individual timed words, and exports the result without uploading the media file.
Step by step
The practical workflow
- 1
Identify the spoken language
Choose an English-only model only when the recording is in English; otherwise use multilingual.
- 2
Start with available memory
Tiny is the safest choice for constrained devices and the fastest first test.
- 3
Evaluate the first transcript
Listen for missed words, names and noisy sections rather than assuming a larger model is always necessary.
- 4
Move up only when useful
Try Base or Small when the expected accuracy gain justifies the extra download, memory and processing time.
Models currently offered
Tiny Q5 is approximately 31 MB, Tiny Q8 approximately 42 MB, Base Q5 approximately 57 MB and Small Q5 approximately 181 MB. English-only Q5 variants are available for Tiny, Base and Small.
These are quantized whisper.cpp model files. Quantization reduces model size and memory requirements with a possible accuracy tradeoff compared with larger or less-quantized files.
Tiny, Base or Small?
Tiny is appropriate for speed, testing and limited hardware. Base is a practical next step when Tiny misses too much. Small is substantially heavier and should be selected deliberately, especially in a browser tab.
- Tiny: shortest download and lowest browser memory pressure
- Base: balanced option for many recordings
- Small: higher resource use for a potentially stronger result
- English variants: intended only for English-language speech
Model size is not the only accuracy factor
Microphone quality, compression, overlapping speakers, accents, background noise and specialist terminology all affect the draft. Even a larger model requires review when the text will be published or relied upon.
Why use local transcription?
FAQ
Common questions
Which Whisper model is fastest?
Tiny is the fastest and smallest option in this application.
Should I always choose Small?
No. Small uses much more memory and processing time. Start smaller and move up when the recording needs it.
What does English-only mean?
The .en variants are designed for English speech and should not be used for multilingual or non-English recordings.
What is Q5 versus Q8?
They are quantization levels. Q5 files are smaller; Q8 retains more numerical precision but has a larger download and memory footprint.
Related guides