app.py

import os
import torch
import torchaudio
import tempfile
from TTS.api import TTS  # Offline TTS
from transformers import (
    SeamlessM4TProcessor,
    SeamlessM4TForSpeechToText,
    SeamlessM4TForSpeechToSpeech,
)
import gradio as gr

# Constants
MODEL_NAME = "facebook/hf-seamless-m4t-medium"
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load processor and models
processor = SeamlessM4TProcessor.from_pretrained(MODEL_NAME)
s2t_model = SeamlessM4TForSpeechToText.from_pretrained(MODEL_NAME).to(device).eval()
s2s_model = SeamlessM4TForSpeechToSpeech.from_pretrained(MODEL_NAME).to(device).eval()

# Load offline TTS model (English-only for now)
tts_engine = TTS(model_name="tts_models/en/ljspeech/tacotron2-DDC", progress_bar=False, gpu=False)

# Main translation function
def translate_from_text(text_input, source_lang, target_lang, auto_detect):
    if not text_input.strip():
        return "Empty input text.", None

    # Step 1: Convert input text to speech using offline TTS
    with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as wav_file:
        tts_engine.tts_to_file(text=text_input, file_path=wav_file.name)
        waveform, sr = torchaudio.load(wav_file.name)

    # Step 2: Resample to 16kHz
    waveform = torchaudio.functional.resample(waveform, sr, 16000)
    src = None if auto_detect else source_lang

    # Step 3: Prepare processor input
    inputs = processor(audios=waveform, src_lang=src, return_tensors="pt").to(device)

    # Step 4: Speech-to-Text
    text_tokens = s2t_model.generate(**inputs, tgt_lang=target_lang)
    translated_text = processor.decode(text_tokens[0].tolist(), skip_special_tokens=True)

    # Step 5: Speech-to-Speech
    speech_waveform = s2s_model.generate(**inputs, tgt_lang=target_lang)[0].cpu().numpy().squeeze()
    translated_audio = (16000, speech_waveform)

    return translated_text, translated_audio

# Gradio Interface
iface = gr.Interface(
    fn=translate_from_text,
    inputs=[
        gr.Textbox(label="Input Text"),
        gr.Textbox(label="Source Language (e.g. eng)"),
        gr.Textbox(label="Target Language (e.g. hin)"),
        gr.Checkbox(label="Auto-detect Source Language")
    ],
    outputs=[
        gr.Textbox(label="Translated Text"),
        gr.Audio(label="Translated Speech", type="numpy")
    ],
    title="iVoice Translate (T2T + T2S → S2T + S2S)"
).queue()

# Launch server
if __name__ == "__main__":
    iface.launch(server_name="0.0.0.0", server_port=int(os.environ.get("PORT", 7860)))