This is an experimental fork of WhisperKit with AI-driven optimizations. It is not intended for production use.

Install

brew install vibekernels/tap/whisperkit-vk-cli

Build from source

swift build

Benchmarking

Build the CLI in release mode:

swift build -c release --product whisperkit-cli

Short audio (~11s)

.build/release/whisperkit-cli transcribe \
  --audio-path Tests/WhisperKitTests/Resources/jfk.wav \
  --model large-v2 \
  --verbose

Long audio (~60s)

.build/release/whisperkit-cli transcribe \
  --audio-path Tests/WhisperKitTests/Resources/ted_60.m4a \
  --model large-v2 \
  --verbose

Comparing models

Run the same file across different models to compare speed and accuracy:

for model in tiny base small large-v2 large-v3; do
  echo "=== $model ==="
  .build/release/whisperkit-cli transcribe \
    --audio-path Tests/WhisperKitTests/Resources/jfk.wav \
    --model $model \
    --verbose
done

Benchmark script

Run all models against short and long audio:

./scripts/benchmark.sh --all

See ./scripts/benchmark.sh --help for options like --audio, --models, and --long.

The --verbose flag prints tokens per second, real-time factor, and speed factor after each transcription.

License

WhisperKit is released under the MIT License.

Upstream

Based on argmaxinc/WhisperKit.