scripts/compare_whisper.py

from pywhispercpp.model import Model
from pathlib import Path
import time
import csv

from silero_vad.utils_vad import languages


def save_csv(file_path, rows):
    with open(file_path, "w", encoding="utf-8") as f:
        writer = csv.writer(f)
        writer.writerows(rows)
        print(f"write csv to {file_path}")

def run_audios_after_vad(models_dir, audio_dir, model_name):
    t0 = time.time()
    model = Model(
                model=model_name,
                models_dir=models_dir,
                print_realtime=False,
                print_progress=False,
                print_timestamps=False,
                translate=False,
                # beam_search=1,
                temperature=0.,
                no_context=True
            )
    print("load model time: ", time.time()-t0)
    rows = [["lang", "file_name", "inference_time", "python_res", "intel_res"]]
    for lang in ["es", "fr", "hi", "it", "ja", "pt"]:
        print("*" * 10, lang, "*"*10)
        for audio in sorted(list((audio_dir/lang).glob("*.wav"))):
            print("Audio name:", audio.name)
            t1 = time.time()
            output = model.transcribe(str(audio), language=lang)
            t = time.time() - t1
            print("Inference time:", t)
            # print(output)
            text = " ".join([a.text for a in output])
            print("Text from Python:", text)
            try:
                with open(audio.with_suffix(".txt"), encoding="utf-8") as f:
                    intel_text = f.read().strip()
            except Exception as e:
                intel_text = ""
                print(f"Error reading Intel text for {audio.name}: {e}")
            print("Text from Intel :", intel_text)
            rows.append([lang, audio.name, t, text, intel_text])
    save_csv("csv/compare_whisper_intel.csv", rows)

def run_long_audios(models_dir, audios_list, model_name):
    t0 = time.time()
    model = Model(
        model=model_name,
        models_dir=models_dir,
        print_realtime=False,
        print_progress=False,
        print_timestamps=False,
        translate=False,
        # beam_search=1,
        temperature=0.,
        no_context=True
    )
    print("load model time: ", time.time() - t0)
    rows = [["file_name", "inference_time", "res_text"]]
    audios = audios_list.read_text().splitlines()
    for audio in audios:
        if not audio:
            rows.append([])
            continue
        lang = Path(audio).name.split('-')[0]
        if lang not in ["es", "fr", "hi", "it", "ja", "pt"]:
            lang = "en"
        print(f"Audio file: {audio}, lang: {lang}")
        t1 = time.time()
        output = model.transcribe(str(audio), language=lang)
        t = time.time() - t1
        print("Inference time:", t)
        # print(output)
        text = " ".join([a.text for a in output])
        print("Text:", text)
        rows.append([audio, t, text])
    save_csv("csv/compare_whisper.csv", rows)
if __name__ == '__main__':
    models_dir = Path("/Users/jeqin/work/code/Translator/python_server/moyoyo_asr_models")
    # model_name = "large-v3-turbo-q5_0"
    model_name = "large-v3-turbo-q8_0"
    # model_name = "small-q8_0"
    # audios_after_vad = Path("/Users/jeqin/work/test/test_yoyotranslator/audios_after_vad/audio2-with-noise")
    audios_list = Path("/Users/jeqin/work/code/TestTranslator/scripts/audios.txt")
    # run_audios_after_vad(models_dir, audios_after_vad, model_name)
    run_long_audios(models_dir, audios_list, model_name)