Aizazayyubi/khowar-whisper-asr

Khowar Automatic Speech Recognition (ASR) Model

This is a low-resource Automatic Speech Recognition (ASR) model for the Khowar language, based on the OpenAI Whisper-base model. It is designed to transcribe spoken Khowar into text accurately and efficiently.

Model Overview

• Language: Khowar (primarily spoken in Chitral, Pakistan)
• Base Model: OpenAI Whisper-base
• Pipeline: Automatic Speech Recognition (ASR)
• Intended Use: Transcription of Khowar audio, research in low-resource ASR, multilingual speech processing
• Tags: ASR, Khowar, Chitrali, low-resource, multilingual

Features

• Supports Khowar speech recognition with reasonable accuracy despite low-resource data.
• Can be integrated into applications for real-time or batch transcription.
• Provides a foundation for building multilingual ASR systems including regional languages of Pakistan.

Model Training

• Data: Collected from local speakers of Khowar, covering a range of accents and contexts.
• Preprocessing: Audio normalized, resampled to 16 kHz.
• Fine-tuning: Whisper-base model fine-tuned on Khowar dataset.
• Evaluation: Model evaluated using Word Error Rate (WER).

Intended Use

• Transcribing Khowar audio recordings.
• Research in multilingual and low-resource ASR.
• Applications requiring integration of Khowar speech-to-text capabilities.

How to Use

You can use the model with Hugging Face Transformers or Whisper pipelines:

from transformers import WhisperForConditionalGeneration, WhisperProcessor

processor = WhisperProcessor.from_pretrained("your-username/khowar-asr")
model = WhisperForConditionalGeneration.from_pretrained("your-username/khowar-asr")

audio_input = processor("path/to/audio.wav", return_tensors="pt").input_features
predicted_ids = model.generate(audio_input)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]

print(transcription)