tried using app.py similar to https://huggingface.co/spaces/openai/whisper/blob/main/app.py

app.py

@@ -1,48 +1,26 @@
+import os
+os.system("pip install git+https://github.com/openai/whisper.git")
 import gradio as gr
-import librosa
-import torch
+import whisper
 
 from transformers import WhisperProcessor, WhisperForConditionalGeneration
 
-processor = WhisperProcessor.from_pretrained("openai/whisper-medium")
-model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-medium")
+model = whisper.load_model("small")
 
-model.config.forced_decoder_ids = None
 
-def process_audio(sampling_rate, waveform):
-    # convert from int16 to floating point
-    waveform = waveform / 32678.0
-
-    # convert to mono if stereo
-    if len(waveform.shape) > 1:
-        waveform = librosa.to_mono(waveform.T)
-
-    # resample to 16 kHz if necessary
-    if sampling_rate != 16000:
-        waveform = librosa.resample(waveform, orig_sr=sampling_rate, target_sr=16000)
-
-    # limit to 30 seconds
-    waveform = waveform[:16000*30]
-
-    # make PyTorch tensor
-    waveform = torch.tensor(waveform)
-    return waveform
-
-def predict(audio, mic_audio=None):
-    # audio = tuple (sample_rate, frames) or (sample_rate, (frames, channels))
-    if mic_audio is not None:
-        sampling_rate, waveform = mic_audio
-    elif audio is not None:
-        sampling_rate, waveform = audio
-    else:
-        return "(please provide audio)"
-
-    waveform = process_audio(sampling_rate, waveform)
-    input_features = processor(waveform, sampling_rate=16000, return_tensors="pt").input_features
-    predicted_ids = model.generate(input_features, max_length=400)
-    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
-    print(transcription)
-    return transcription[0]
+def inference(audio):
+    audio = whisper.load_audio(audio)
+    audio = whisper.pad_or_trim(audio)
+    
+    mel = whisper.log_mel_spectrogram(audio).to(model.device)
+    
+    _, probs = model.detect_language(mel)
+    
+    options = whisper.DecodingOptions(fp16 = False)
+    result = whisper.decode(model, mel, options)
+    
+    print(result.text)
+    return result.text, gr.update(visible=True), gr.update(visible=True), gr.update(visible=True)
 
 
 title = "Demo for Whisper -> Something -> XLS-R"
@@ -53,7 +31,7 @@ being passed into the model. The output is the text transcription of the audio.
 """
 
 gr.Interface(
-    fn=predict,
+    fn=inference,
     inputs=[
         gr.Audio(label="Upload Speech", source="upload", type="numpy"),
         gr.Audio(label="Record Speech", source="microphone", type="numpy"),