app.py

import gradio as gr
import torch
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import soundfile as sf
import numpy as np
from scipy import signal
import os

# Set the cache directory to a writable location
cache_dir = "/tmp/.cache"
os.environ["TRANSFORMERS_CACHE"] = cache_dir
os.environ["HF_DATASETS_CACHE"] = cache_dir
os.environ["TORCH_HOME"] = cache_dir  # Set PyTorch cache directory

# Ensure the cache directory exists and is writable
os.makedirs(cache_dir, exist_ok=True)

# Load the base Whisper model and processor
def load_model():
    print("Loading base Whisper model and processor...")
    processor = WhisperProcessor.from_pretrained("openai/whisper-large-v3")
    model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-large-v3")
    
    # Load the fine-tuned weights from the Macedonian-ASR repository
    print("Loading fine-tuned weights from Macedonian-ASR...")
    model.load_state_dict(torch.hub.load_state_dict_from_url(
        "https://huggingface.co/Macedonian-ASR/whisper-large-v3-macedonian-asr/resolve/main/pytorch_model.bin",
        map_location="cpu",
        model_dir=cache_dir  # Save downloaded weights to the writable cache directory
    ))
    print("✓ Model and processor loaded successfully!")
    return processor, model

processor, model = load_model()

def process_audio(audio_path):
    # Load and resample to 16kHz using scipy
    waveform, sr = sf.read(audio_path)
    if len(waveform.shape) > 1:  # Convert stereo to mono
        waveform = waveform.mean(axis=1)
    if sr != 16000:  # Resample if necessary
        num_samples = int(len(waveform) * 16000 / sr)
        waveform = signal.resample(waveform, num_samples)
    
    # Process the audio
    inputs = processor(waveform, sampling_rate=16000, return_tensors="pt")
    predicted_ids = model.generate(**inputs, language="mk")
    transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)[0]
    return transcription

# Gradio interface
demo = gr.Interface(
    fn=process_audio,
    inputs=gr.Audio(sources=["microphone", "upload"], type="filepath"),
    outputs="text",
    title="Македонско препознавање на говор / Macedonian Speech Recognition",
    description="Качете аудио или користете микрофон за транскрипција на македонски говор / Upload audio or use microphone to transcribe Macedonian speech"
)

if __name__ == "__main__":
    demo.launch()