Update README.md

README.md

@@ -2,60 +2,45 @@
 license: apache-2.0
 datasets:
 - ARTPARK-IISc/Vaani
-- google/fleurs
 language:
-- kn
+- hi
 base_model:
 - openai/whisper-small
 pipeline_tag: automatic-speech-recognition
 ---
 
-```Python
-
-import torch
-from transformers import WhisperForConditionalGeneration, WhisperProcessor, WhisperTokenizer,WhisperFeatureExtractor
-import soundfile as sf
-
-
-model="ARTPARK-IISc/whisper-small-vaani-kannada"
-
-# Load tokenizer and feature extractor individually
-feature_extractor = WhisperFeatureExtractor.from_pretrained(model)
-tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-small", language="Kannada", task="transcribe")
 
+# Whisper-small-vaani-kannada
 
-# Create the processor manually
-processor = WhisperProcessor(feature_extractor=feature_extractor, tokenizer=tokenizer)
+This is a fine-tuned version of [OpenAI's Whisper-Small](https://huggingface.co/openai/whisper-small), trained on Kannada speech from multiple datasets.
 
-# Load and preprocess the audio file
-audio_file_path = "Sample_Audio.wav"  # replace with your audio file path
-
-
-device = "cuda" if torch.cuda.is_available() else "cpu"
-
-# Load the processor and model
-model = WhisperForConditionalGeneration.from_pretrained(model).to(device)
+# Usage
+This can be used with the pipeline function from the Transformers module.
+```python
 
+import torch
+from transformers import pipeline
 
-# load audio
-audio_data, sample_rate = sf.read(audio_file_path)
-# Ensure the audio is 16kHz (Whisper expects 16kHz audio)
-if sample_rate != 16000:
-    import torchaudio
-    resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=16000)
-    audio_data = resampler(torch.tensor(audio_data).unsqueeze(0)).squeeze().numpy()
+audio = "path to the audio file to be transcribed"
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
+modelTags="ARTPARK-IISc/whisper-small-vaani-kannada"
+transcribe = pipeline(task="automatic-speech-recognition", model=modelTags, chunk_length_s=30, device=device)
+transcribe.model.config.forced_decoder_ids = transcribe.tokenizer.get_decoder_prompt_ids(language="ka", task="transcribe")
 
+print('Transcription: ', transcribe(audio)["text"])
 
-# Use the processor to prepare the input features
-input_features = processor(audio_data, sampling_rate=16000, return_tensors="pt").input_features.to(device)
+```
+# Training and Evaluation
 
-# Generate transcription (disable gradient calculation during inference)
-with torch.no_grad():
-    predicted_ids = model.generate(input_features)
+The models has finetuned using folllowing dataset [Vaani](https://huggingface.co/datasets/ARTPARK-IISc/Vaani) , [Fleurs](https://huggingface.co/datasets/google/fleurs),[IndicTTS](https://huggingface.co/datasets/SPRINGLab/IndicTTS-Hindi)
 
-# Decode the generated IDs into human-readable text
-transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
 
-print(transcription)
+The performance of the model was evaluated using multiple datasets, and the evaluation results are provided below.
 
-```
+| Dataset | WER | 
+| :---:   | :---: | 
+| Fleurs | 29.16   | 
+| IndicTTS | 15.27   | 
+| Kathbath | 33.94   | 
+| Kathbath Noisy| 38.46  | 
+| Vaani  | 69.78  |