app.py

import gradio as gr
import os
import torch
import soundfile as sf
import torchaudio
from scipy.io.wavfile import write
from transformers import VitsProcessor, VitsForConditionalGeneration
from speechbrain.pretrained import EncoderClassifier, EncoderDecoderASR

# Ensure the output directory exists
os.makedirs("output_audio", exist_ok=True)

# Load the Facebook MMS TTS model and processor
tts_processor = VitsProcessor.from_pretrained("facebook/mms-tts-eng")
tts_model = VitsForConditionalGeneration.from_pretrained("facebook/mms-tts-eng")

# SpeechBrain ASR Model for Speech to Text
asr_model = EncoderDecoderASR.from_hparams(source="speechbrain/asr-conformer-transformerlm-librispeech", savedir="models/asr")

def speech_to_text(input_audio):
    sig, sr = torchaudio.load(input_audio)
    if sr != 16000:
        sig = torchaudio.transforms.Resample(orig_freq=sr, new_freq=16000)(sig)
    transcription = asr_model.transcribe_file(input_audio)
    return transcription

def text_to_speech(text):
    inputs = tts_processor(text, return_tensors="pt")
    with torch.no_grad():
        generated = tts_model.generate(**inputs)
        waveform = generated.audio.squeeze().cpu().numpy()

    output_path = "output_audio/text_to_speech.wav"
    sf.write(output_path, waveform, 22050)
    return output_path

def speech_to_speech(input_audio, target_text):
    # Speech to Text
    transcription = speech_to_text(input_audio)
    
    # Text to Speech with Facebook MMS TTS
    return text_to_speech(target_text)

iface = gr.Interface(
    fn={
        "Speech to Text": speech_to_text,
        "Text to Speech": text_to_speech,
        "Speech to Speech": speech_to_speech
    },
    inputs={
        "Speech to Text": gr.inputs.Audio(source="upload", type="file"),
        "Text to Speech": gr.inputs.Textbox(label="Text"),
        "Speech to Speech": [gr.inputs.Audio(source="upload", type="file"), gr.inputs.Textbox(label="Target Text")]
    },
    outputs={
        "Speech to Text": gr.outputs.Textbox(label="Transcription"),
        "Text to Speech": gr.outputs.Audio(type="file", label="Synthesized Speech"),
        "Speech to Speech": gr.outputs.Audio(type="file", label="Synthesized Speech")
    },
    title="Speech Processing App",
    description="This app uses SpeechBrain for speech to text and Facebook's MMS for text to speech.",
    layout="vertical"
)

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