app.py

import torch
import torchaudio
import soundfile as sf
import gradio as gr
from transformers import SeamlessM4TProcessor, SeamlessM4TModel

# Load model and processor
HF_MODEL_ID = "facebook/hf-seamless-m4t-medium"
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
processor = SeamlessM4TProcessor.from_pretrained(HF_MODEL_ID)
model = SeamlessM4TModel.from_pretrained(HF_MODEL_ID).to(device).eval()

def translate_backend(audio_path, src_lang, tgt_lang):
    try:
        waveform, sr = sf.read(audio_path)
        if len(waveform.shape) > 1:
            waveform = waveform.mean(axis=1)
        waveform = waveform.astype("float32")

        if sr != 16000:
            waveform_tensor = torch.tensor(waveform).unsqueeze(0)
            resampler = torchaudio.transforms.Resample(orig_freq=sr, new_freq=16000)
            waveform = resampler(waveform_tensor).squeeze(0).numpy()
            sr = 16000

        inputs = processor(audios=waveform, sampling_rate=sr, return_tensors="pt", src_lang=src_lang).to(device)
        with torch.no_grad():
            output = model.generate(**inputs, tgt_lang=tgt_lang, generate_speech=False)
        translated_text = processor.batch_decode(output.sequences, skip_special_tokens=True)[0]

        return translated_text
    except Exception as e:
        return f"❌ Error: {str(e)}"

# Standard Gradio launch (no API flag)
iface = gr.Interface(
    fn=translate_backend,
    inputs=[
        gr.Audio(type="filepath", label="🎤 Audio"),
        gr.Text(label="Source Language"),
        gr.Text(label="Target Language")
    ],
    outputs="text",
    title="Kalpani iVoice Translate",
    allow_flagging="never"
)

if __name__ == "__main__":
    iface.launch()  # No 'api=True'