ivrit-faster-whisper-turbo-d4

This model is a conversion of the ivrit-ai/whisper-large-v3-turbo-d4-p1-take2 model to the Faster-Whisper format, offering significantly faster inference times.

Model Overview

• Base Model: ivrit-ai/whisper-large-v3-turbo-d4-p1-take2
• Converted to: Faster-Whisper (for faster ASR with minimal performance loss)
• Language: Hebrew (he)
• Quantization: Float32

All credits go to ivrit-ai for developing the original Whisper model.

How to Use the Model

To use the model in your projects, follow the steps below to load and transcribe audio:

# Import the Faster Whisper module
import faster_whisper

# Load the model from Hugging Face
model = faster_whisper.WhisperModel("israelisraeli/ivrit-faster-whisper-turbo-d4", device="cuda")

# Transcribe the audio file to JSON
segs, _ = model.transcribe("AUDIOFILE_efiTheTigger.mp3", language="he")

# Format the output into a list of dictionaries with timestamps and text
transcribed_segments_with_timestamps = [
    {"start": s.start, "end": s.end, "text": s.text} for s in segs
]

import json

# Save the result to a JSON file
with open("transcribed_segments_with_timestamps.json", "w", encoding="utf-8") as json_file:
    json.dump(
        transcribed_segments_with_timestamps, json_file, ensure_ascii=False, indent=4
    )

print("Transcription saved to transcribed_segments_with_timestamps.json")

Conversion process

Tokenizer Conversion

from transformers import AutoTokenizer

# Load the tokenizer from the original Whisper model files
tokenizer_directory = "path_to_whisper_model_files"
tokenizer = AutoTokenizer.from_pretrained(tokenizer_directory)

# Save the tokenizer into a single JSON file
tokenizer.save_pretrained("path_to_save_directory", legacy_format=False)

Model Conversion to Faster-Whisper

To convert the original ivrit-ai/whisper-large-v3-turbo-d4-p1-take2 model to the Faster-Whisper format, i used the CTranslate2 library. The following command was used for the conversion:

ct2-transformers-converter \
  --model ./whisper-large-v3-turbo-d4-p1-take2 \
  --output_dir ./ivrit-faster-whisper-turbo-d4 \
  --copy_files tokenizer.json preprocessor_config.json \