| --- |
| 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) |
| ```python |
| 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 |
| |
| ```bash |
| whisper-large-v3-turbo-ct2/ |
| │ |
| ├── config.json |
| ├── model.bin |
| ├── preprocessor_config.json |
| ├── tokenizer.json |
| └── vocabulary.json |
| ``` |
