Team Ai
Apppublic

Xenobd/whisper.cpp

sourceHugging Faceupdated 10mo agoView on Hugging Face
0likes
bench-wts.sh71 linesDownload Raw Back to scripts
1# Benchmark word-level timestamps for different models2#3# This script takes two arguments4# - an audio file5# - [optional] path to a font file6 7# I'm using "/usr/share/fonts/truetype/freefont/FreeMono.ttf" on Ubuntu8 9if [ -z "$1" ]; then10    echo "Usage: $0 <audio file> [font file]"11    exit 112fi13 14#TODO: Make this a command line parameter15#models="base small large"16#models="tiny.en tiny base.en base small.en small medium.en medium large-v1 large"17models="tiny.en base.en small.en medium.en large"18 19DURATION=$(ffprobe -i $1 -show_entries format=duration -v quiet -of csv="p=0")20DURATION=$(printf "%.2f" $DURATION)21echo "Input file duration: ${DURATION}s"22 23for model in $models; do24    echo "Running $model"25    COMMAND="./build/bin/whisper-cli -m models/ggml-$model.bin -owts -f $1 -of $1.$model"26 27    if [ ! -z "$2" ]; then28        COMMAND="$COMMAND -fp $2"29    fi30    #TODO: Surface errors better31    # TIMEFMT is for zsh, TIMEFORMAT is for bash32    EXECTIME=$({ TIMEFMT="%E";TIMEFORMAT=%E; time $COMMAND >/dev/null 2>&1; } 2>&1)33 34    # Slightly different formats between zsh and bash35    if [ "${EXECTIME: -1}" == "s" ]; then36        EXECTIME=${EXECTIME::-1}37    fi38 39    RATIO=$(echo "$DURATION / $EXECTIME" | bc -l)40    RATIO=$(printf "%.2f" $RATIO)41 42    echo "Execution time: ${EXECTIME}s (${RATIO}x realtime)"43 44    # If the file already exists, delete it45    if [ -f $1.mp4 ]; then46        rm $1.mp447    fi48 49    bash $1.$model.wts >/dev/null 2>&150    mv $1.mp4 $1.$model.mp451 52    ffmpeg -y -f lavfi -i color=c=black:s=1200x50:d=$DURATION -vf "drawtext=fontfile=$2:fontsize=36:x=10:y=(h-text_h)/2:text='ggml-$model - ${EXECTIME}s (${RATIO}x realtime)':fontcolor=lightgrey" $1.$model.info.mp4 >/dev/null 2>&153done54 55COMMAND="ffmpeg -y"56for model in $models; do57    COMMAND="$COMMAND -i $1.$model.info.mp4 -i $1.$model.mp4"58done59COMMAND="$COMMAND -filter_complex \""60COUNT=061for model in $models; do62    COMMAND="$COMMAND[${COUNT}:v][$(($COUNT+1)):v]"63    COUNT=$((COUNT+2))64done65COMMAND="$COMMAND vstack=inputs=${COUNT}[v]\" -map \"[v]\" -map 1:a $1.all.mp4 >/dev/null 2>&1"66 67echo $COMMAND68 69# Run the command70eval $COMMAND71