Faster Whisper transcribe script

README.md

Faster Whisper transcribe script

This is a helper script to make the process of transcribing with guillaumekln/faster-whisper model easier.

Installation

1. Create a venv python -m venv venv
2. Activate the venv source venv/bin/active
3. Install faster-whisper from PyPI pip install faster-whisper
4. Download the faster-whisper-large-v2 model from Hugging Face
5. Move the Hugging Face large-v2 model contents into faster-whisper-large-v2-ct2 folder
6. Run a transcription job: python transcribe.py audio.mp3

transcribe.py

import argparse
from datetime import datetime, timedelta
from pathlib import Path
from typing import Optional

from faster_whisper import WhisperModel


# model_size = "large-v2"  # Will download faster-whisper-large-v2 from Hugging face
model_size = "faster-whisper-large-v2-ct2"  # Load the local large-v2 model

# Run on GPU with FP16
# model = WhisperModel(model_size, device="cuda", compute_type="float16")

# or run on GPU with INT8
# model = WhisperModel(model_size, device="cuda", compute_type="int8_float16")
# or run on CPU with INT8
# model = WhisperModel(model_size, device="cpu", compute_type="int8")


parser = argparse.ArgumentParser(
    usage="%(prog)s [audio file]",
    description="Transcribe the given audio file using faster-whisper-large-v2 model",
    allow_abbrev=False,
)
parser.add_argument("audio_file", type=str)
parser.add_argument(
    "lang",
    type=str,
    nargs="?",
    default=None,
    help="Specify the audio language. Will auto-detect if not specified",
)
args = parser.parse_args()


def srt_format_timestamp(seconds: float) -> str:
    """
    Taken from: https://github.com/openai/whisper/discussions/98#discussioncomment-3726175
    """
    assert seconds >= 0, "non-negative timestamp expected"
    milliseconds = round(seconds * 1000.0)

    hours = milliseconds // 3_600_000
    milliseconds -= hours * 3_600_000

    minutes = milliseconds // 60_000
    milliseconds -= minutes * 60_000

    seconds = milliseconds // 1_000
    milliseconds -= seconds * 1_000

    return (f"{hours:02d}:") + f"{minutes:02d}:{seconds:02d},{milliseconds:03d}"


def transcribe_audio(audio_path: str, audio_lang: Optional[str]) -> None:
    model = WhisperModel(model_size, device="cpu", compute_type="int8")

    segments, info = model.transcribe(audio_path, language=audio_lang, beam_size=5)

    if audio_lang is None:
        print(f"Detected language '{info.language}' with probability {info.language_probability}")

    now_time = datetime.now().strftime("%Y-%m-%d_%H-%M")
    file_name_no_suffix = Path(audio_path).with_suffix("").name
    srt_file_name = Path(f"{file_name_no_suffix}_{now_time}.srt").name

    transcription = ""
    for segment in segments:
        startTime = srt_format_timestamp(segment.start)
        endTime = srt_format_timestamp(segment.end)
        text_srt = (f"{segment.id}\n{startTime} --> {endTime}\n{segment.text.strip()}\n\n")
        text_console = (f"[{segment.id} - {startTime} --> {endTime}] {segment.text.strip()}")

        print(text_console)
        transcription += text_srt

    Path(srt_file_name).write_text(transcription)


transcribe_audio(args.audio_file, args.lang)