whisper-Large-v3-hindi-ct2

Whisper Large v3 Hindi (CTranslate2 Conversion)

Repository: whisper-Large-v3-hindi-ct2
Original Model: quinnb/whisper-Large-v3-hindi
Converted with: CTranslate2
Intended Use: Automatic Speech Recognition (ASR) for Hindi

---

Model Description

This repository hosts a CTranslate2-converted version of the quinnb/whisper-Large-v3-hindi model. The original model is based on OpenAI’s Whisper architecture, specialized for Hindi speech recognition. Converting it to CTranslate2 format can significantly speed up inference and reduce memory usage, especially when used with libraries such as faster-whisper.

Key Benefits:

• Reduced Inference Latency: CTranslate2 uses optimized kernels for quick speech-to-text transcription.
• Lower Memory Footprint: Quantization (float16, int8_float16, etc.) and efficient GPU usage.
• Easy Integration: Compatible with faster-whisper, enabling direct usage in Python.

---

Usage

Installation

You’ll need:

• Python 3.8+
• CTranslate2
• faster-whisper

Example Installation:

pip install faster-whisper ctranslate2

Loading the Model

from faster_whisper import WhisperModel

model_path = "akanshSirohi/whisper-Large-v3-hindi-ct2"

model = WhisperModel(model_path, device="cuda")  # use "cpu" if no GPU

audio_file = "your_audio.wav"
segments, info = model.transcribe(audio_file)

print("Detected language:", info.language)
for segment in segments:
    print(f"[{segment.start:.2f}s -> {segment.end:.2f}s] {segment.text}")

---

Intended Use Cases and Limitations

1. Hindi Speech Recognition: Designed for transcribing Hindi audio into text.
2. Transcription and Subtitling: Ideal for video subtitling, podcast transcription, or voice notes.

Not Recommended For:

• Non-Hindi languages.
• Real-time streaming speech (though it might still be possible with enough optimization).

Potential Risks & Limitations:

• Accuracy may be lower for dialects or code-mixed speech (Hindi + other languages).
• Biases in training data might affect performance for certain accents or topics.

---

Training and Sources

• Original Weights: From quinnb/whisper-Large-v3-hindi.
• Base Architecture: OpenAI’s Whisper large v3.

See the original model card for full training details.

---

License

This model follows the same license as OpenAI Whisper. Generally, it’s under the OpenAI Whisper License. Please ensure compliance when using or distributing outputs from the model.

---

Citation

If you find this model or its outputs useful for your research or applications, please cite accordingly:

@misc{openaiwhisper,
  title={Whisper: Robust Speech Recognition via Large-Scale Weak Supervision},
  author={OpenAI},
  year={2022},
  note={\url{https://github.com/openai/whisper}}
}

And also mention CTranslate2 in your work for the conversion/optimization framework.

---

Acknowledgements

• OpenAI for developing Whisper.
• Guillaume Klein for faster-whisper.
• CTranslate2 for efficient inference engine.

---

Model Card Authors

• Akansh Sirohi (@akanshSirohi)

If you have questions, issues, or suggestions, feel free to open a discussion or pull request!