transcribe.py

from faster_whisper import WhisperModel
import torch
import gc
import json

gc.collect()
torch.cuda.empty_cache()

model = WhisperModel("medium", device="cuda", compute_type="int8_float16")


def start_transcribe(progress):
    _, speaker_groups = load_groups_json()

    subtitle_txt = []
    for speaker, _ in zip(speaker_groups, progress.tqdm(speaker_groups, desc="Transcribing")):
        # Transcribe and save temp file
        audiof = f"{speaker}.wav"
        print(f"Loading {audiof}")
        segments, _ = model.transcribe(
            audio=audiof, language='id', word_timestamps=True)
        segments_list = list(segments)

        text_list_to_print = []
        for segment in segments_list:
            start = timeStr(segment.start)
            end = timeStr(segment.end)
            name = str(speaker)[:10]
            text = segment.text
            subtitle_txt.append(
                f"{len(subtitle_txt) + 1}\n{start} --> {end}\n[{name}] {text}\n\n")
            # Appending text for each segment to print
            text_list_to_print.append(text)

        # Print full text for each speaker turn
        text = "\n".join(text_list_to_print)
        print(text)
        # Append to complete transcribe file
        with open("transcribe.txt", "a") as file:
            file.write(f"[{name}] {text}\n")

    # Appending subtitle txt for each segment
    with open("subtitle.srt", "w") as file:
        file.writelines(subtitle_txt)
    return ["transcribe.txt", "subtitle.srt"]


def timeStr(t):
    return '{0:02d}:{1:02d}:{2:06.2f}'.format(round(t // 3600),
                                              round(t % 3600 // 60),
                                              t % 60)


def load_groups_json():
    with open("sample_groups.json", "r") as json_file_sample:
        sample_groups_list: list = json.load(json_file_sample)
    with open("speaker_groups.json", "r") as json_file_speaker:
        speaker_groups_dict: dict = json.load(json_file_speaker)
    return sample_groups_list, speaker_groups_dict