#uvicorn app:app --host 0.0.0.0 --port 8000 --reload
from fastapi import FastAPI, UploadFile, File
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from fastapi.responses import HTMLResponse
import librosa
import io
import re
html_tag_remover = re.compile(r'<[^>]+>')
def remove_tags(text):
return html_tag_remover.sub('', text)
app = FastAPI()
processor = WhisperProcessor.from_pretrained("openai/whisper-medium")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-medium")
model.config.forced_decoder_ids = None
chunk_duration = 30
overlap_duration = 5
@app.get("/")
def read_root():
html_form = """
<html>
<body>
<h2>ASR Transcription</h2>
<form action="/transcribe" method="post" enctype="multipart/form-data">
<label for="audio_file">Upload an audio file (MP3 or WAV):</label>
<input type="file" id="audio_file" name="audio_file" accept=".mp3, .wav" required><br><br>
<input type="submit" value="Transcribe">
</form>
</body>
</html>
"""
return HTMLResponse(content=html_form, status_code=200)
@app.post("/transcribe")
async def transcribe_audio(audio_file: UploadFile):
audio_data = await audio_file.read()
audio_data, _ = librosa.load(io.BytesIO(audio_data), sr=16000)
transcription = []
start = 0
while start < len(audio_data):
end = start + chunk_duration * 16000
audio_chunk = audio_data[start:end]
input_features = processor(audio_chunk.tolist(), return_tensors="pt").input_features
predicted_ids = model.generate(input_features, max_length=1000)
chunk_transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
transcription.extend(chunk_transcription)
start = end - overlap_duration * 16000
final_transcription = " ".join(transcription)
final_transcription = remove_tags(final_transcription)
return final_transcription