app.py

import torch
import torchaudio  # ✅ Added torchaudio to handle audio resampling
import gradio as gr
import time  
import numpy as np
import scipy.io.wavfile
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline

# ✅ 1️⃣ Force Model to Run on CPU
device = "cpu"
torch_dtype = torch.float32  # Use CPU-friendly float type
MODEL_NAME = "openai/whisper-tiny"  # ✅ Switched to smallest model for fastest performance

# ✅ 2️⃣ Load Whisper Tiny Model on CPU (Removed `low_cpu_mem_usage=True`)
model = AutoModelForSpeechSeq2Seq.from_pretrained(
    MODEL_NAME, torch_dtype=torch_dtype, use_safetensors=True  # ✅ Removed low_cpu_mem_usage
)
model.to(device)

# ✅ 3️⃣ Load Processor & Pipeline
processor = AutoProcessor.from_pretrained(MODEL_NAME)

pipe = pipeline(
    task="automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    chunk_length_s=2,  # ✅ Process in 2-second chunks for ultra-low latency
    torch_dtype=torch_dtype,
    device=device,
    sampling_rate=16000,  # ✅ Explicitly set sampling rate to avoid resampling issues
)

# ✅ 4️⃣ Real-Time Streaming Transcription (Microphone)
def stream_transcribe(stream, new_chunk):
    start_time = time.time()
    try:
        sr, y = new_chunk

        # ✅ Convert stereo to mono
        if y.ndim > 1:
            y = y.mean(axis=1)
            
        y = y.astype(np.float32)
        y /= np.max(np.abs(y))

        # ✅ Resample audio