metadata
license: apache-2.0
language:
- fa
base_model:
- SadeghK/whisper-large-v3-turbo
tags:
- ASR
- Persian
- Farsi
Whisper Large V3 Turbo (CTranslate2) — Optimized for Faster-Whisper & WhisperX
This repository contains a CTranslate2 (CT2) optimized version of the whisper-large-v3-turbo model finetuned by SadeghK.
It is designed for high-speed inference, low-latency ASR, and full WhisperX compatibility (ASR + alignment + diarization).
🚀 Model Overview
This is a converted version of the original SadeghK/whisper-large-v3-turbo model into CTranslate2 format, which enables:
- ✔ Faster inference (up to 4× vs PyTorch)
- ✔ Lower memory usage (supports float16 / int8 / int8_float16)
- ✔ Full compatibility with faster-whisper
- ✔ Full compatibility with WhisperX for:
- ASR transcription
- Word-level alignment
- (optional) speaker diarization
All weights in this repository are ready-to-use, no additional conversion required.
🔬 Usage with WhisperX (ASR + alignment)
import whisperx
device = "cuda"
# ASR
asr_model = whisperx.load_model(
"SadeghK/whisper-large-v3-turbo-ct2",
device=device,
compute_type="float16"
)
result = asr_model.transcribe("audio.wav")
# Alignment (example for Persian)
align_model, metadata = whisperx.load_align_model("fa", device)
aligned = whisperx.align(result["segments"], align_model, metadata, "audio.wav", device)
📁 Repository Structure
whisper-large-v3-turbo-ct2/
│
├── config.json
├── model.bin
├── preprocessor_config.json
├── tokenizer.json
└── vocabulary.json