Python file to compare the inference times between whisper.cpp Python bindings and faster-whisper

test_transcribe.py

import argparse
import whispercpp as w
from faster_whisper import WhisperModel, decode_audio

import time
from collections.abc import Iterator
from contextlib import contextmanager

@contextmanager
def time_it() -> Iterator[None]:
    tic: float = time.perf_counter()
    try:
        yield
    finally:
        toc: float = time.perf_counter()
        print(f"Computation time = {1000*(toc - tic):.3f}ms")

def transcribe_file_cpp(file_path, model_name):
    transcriber = w.Whisper.from_pretrained(model_name)
    res = transcriber.transcribe_from_file(file_path)
    return res

def transcribe_file_fwhisper(file_path, model_name):
    model = WhisperModel(model_name)
    segments, info = model.transcribe(file_path, word_timestamps=True)
    return segments

if __name__ == "__main__":
    parser = argparse.ArgumentParser()
    parser.add_argument("--file_path", default="jfk.wav", help="Path to the audio file", nargs='?')
    parser.add_argument("--model_name", default="tiny.en", choices=list(w.utils.MODELS_URL), help="Name of the model to use for transcription")
    args = parser.parse_args()

    with time_it():
        res = transcribe_file_cpp(args.file_path, args.model_name)
    print(res)

    with time_it():
        res = transcribe_file_fwhisper(args.file_path, args.model_name)
    for segment in res:
        print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))