app.py

import torch
import gradio as gr
import torchaudio
import time
from datetime import datetime
import numpy as np  # Add this import for handling numpy arrays
from transformers import pipeline
from tortoise.api import TextToSpeech
from tortoise.utils.text import split_and_recombine_text
from tortoise.utils.audio import load_audio, load_voice, load_voices

# STT Initialization
model_id = "openai/whisper-tiny"
pipe = pipeline("automatic-speech-recognition", model=model_id)

# TTS Initialization
VOICE_OPTIONS = [
    "indian_f_1", "indian_F_2", "indian_F_3",
    "indian_M_1", "indian_M_2", "indian_M_3"
]
tts = TextToSpeech(kv_cache=True, use_deepspeed=True, half=True)

def convert_audio(filepath, voice="indian_F_1"):
    # Transcribe audio to text using STT
    
    )
    transcribed_text = transcription_output["text"]
    
    # Use the transcribed text for TTS
    texts = split_and_recombine_text(transcribed_text)
    voice_samples, conditioning_latents = load_voice(voice)
    audio_frames = []
    for text in texts:
        for audio_frame in tts.tts_with_preset(
            text,
            voice_samples=voice_samples,
            
            k=1
        ):
            audio_frames.append(audio_frame.cpu().detach().numpy())

    # Joining the audio frames for output using numpy's concatenate
    final_audio = np.concatenate(audio_frames, axis=0)
    

interface = gr.Interface(
    fn=convert_audio,
    inputs=[
        gr.Audio(source="upload", type="filepath"),
        gr.Dropdown(VOICE_OPTIONS, value="indian_f_1", label="Select voice:", type="value")
    ],
    outputs=gr.Audio(label="streaming audio:", streaming=True, autoplay=True),
    title="STT to TTS",
    description="Convert spoken words into a different voice"
)

interface.launch()