Benchmark Setup of whisper.cpp vs faster-whisper

[wtlow003]

I understand from the README.md that faster-whisper is supposedly "faster" than whisper.cpp.

However, from my own setup – I couldn't get faster-whisper to run as fast as whisper.cpp. I am using implementations of faster-whisper from stable-ts and whisper-ctranslate2.

For reference, I am using Apple M2 Max (14-inch), some commands I ran for whisper-cpp:

with mps:

"$WHISPER_BIN_PATH/main" -m "$WHISPER_BIN_PATH/models/ggml-medium.bin" \
            -f "$saved_dir/audio.wav" \
            -ml 47 \
            -osrt \
            -of "$saved_dir/subs" \
            -t 8 \
            -p 1 \
            --split-on-word

without mps:

"$WHISPER_BIN_PATH/main" -m "$WHISPER_BIN_PATH/models/ggml-medium.bin" \
            -f "$saved_dir/audio.wav" \
            -ml 47 \
            -osrt \
            -of "$saved_dir/subs" \
            -t 8 \
            -p 1 \
            --split-on-word
            -ng

Commands via whisper-ctranslate2:

whisper-ctranslate2 ~/code/auto-subtitles/output/pytorch-in-100-seconds/audio.wav --model medium --language en --print_color TRUE --compute_type float32 --threads 8 --max_line_width 47 --word_timestamps True