PhoWhisper-large-ct2

Fork notice: this is a fork of kiendt/PhoWhisper-large-ct2 (MIT license). Only the model card metadata was changed — library_name: ctranslate2 and tags were added so HF-tag-based discovery/filtering tools (e.g. Speaches) recognize this as a valid CTranslate2 ASR checkpoint. The weights are unmodified, byte-identical to the original repo.

This repository contains the PhoWhisper-large model converted to use CTranslate2 for faster inference. This allows for significant performance improvements, especially on CPU.

Usage

1. Installation: Ensure you have the necessary libraries installed:

   pip install transformers ctranslate2 faster-whisper
   

2. Conversion (only needed once): This step converts the original Hugging Face model to the CTranslate2 format.

   ct2-transformers-converter --model vinai/PhoWhisper-large --output_dir PhoWhisper-large-ct2 --copy_files tokenizer_config.json --quantization float16
   

3. Transcription:

   import os
   from faster_whisper import WhisperModel
   
   model_size = "kiendt/PhoWhisper-large-ct2"
   # 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")
   
   segments, info = model.transcribe("audio.wav", beam_size=5) # Replace audio.wav with your audio file
   
   print("Detected language '%s' with probability %f" % (info.language, info.language_probability))
   
   for segment in segments:
       print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
   

Model Details

• Based on the vinai/PhoWhisper-large model.
• Converted using ct2-transformers-converter.
• Optimized for faster inference with CTranslate2.

Contributing

Contributions are welcome! Please open an issue or submit a pull request.

License

MIT