app.py

import gradio as gr
import torch
import librosa
import numpy as np
from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel

print("Loading model...")
processor = WhisperProcessor.from_pretrained("openai/whisper-large-v3")
base_model = WhisperForConditionalGeneration.from_pretrained(
    "openai/whisper-large-v3",
    torch_dtype=torch.float32,
    low_cpu_mem_usage=True
)
model = PeftModel.from_pretrained(base_model, "Dougsworth/caribbean-whisper-asr")
print("Model loaded!")

def enhance_phone_audio(audio):
    """Gentle enhancement for phone audio - just normalize, no filtering"""
    # Remove DC offset
    audio = audio - np.mean(audio)
    
    # Normalize to [-1, 1]
    max_val = np.max(np.abs(audio))
    if max_val > 0:
        audio = audio / max_val
    
    # Gentle compression to boost quiet parts
    audio = np.sign(audio) * np.power(np.abs(audio), 0.9)
    
    # Normalize again
    max_val = np.max(np.abs(audio))
    if max_val > 0:
        audio = audio / max_val
    
    return audio.astype(np.float32)

def transcribe(audio_path):
    if audio_path is None:
        return "Please upload an audio file."
    
    audio, sr = librosa.load(audio_path, sr=16000)
    
    # Apply gentle enhancement
    audio = enhance_phone_audio(audio)
    
    inputs = processor(audio, sampling_rate=16000, return_tensors="pt")
    
    with torch.no_grad():
        predicted_ids = model.generate(inputs.input_features, language="en", task="transcribe")
    
    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
    return transcription

demo = gr.Interface(
    fn=transcribe,
    inputs=gr.Audio(type="filepath", label="Upload Voice Note"),
    outputs=gr.Textbox(label="Job Listing Transcription"),
    title="Linkup - Caribbean Speech to Text",
    description="Upload a voice note describing a job and get it transcribed. Built for Caribbean accents.",
    examples=[]
)
demo.launch()