app.py

import os
import gradio as gr
import torch
import soundfile as sf
from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC
import librosa
from TTS.api import TTS
from TTS.utils.manage import ModelManager

# Agreeing to Coqui TTS terms of service and setting up environment variables
os.environ["COQUI_TOS_AGREED"] = "1"
os.makedirs("output_audio", exist_ok=True)

# Initialize ASR model
asr_processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-large-960h")
asr_model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-large-960h")
asr_model.eval()

# Dynamically list and select TTS model
tts_manager = ModelManager()
model_name = "tts_models/multilingual/multi-dataset/xtts_v2"
tts = TTS(model_name, gpu=False)

def resample_audio(input_audio_path, target_sr=16000):
    waveform, sr = sf.read(input_audio_path)
    if sr != target_sr:
        waveform = librosa.resample(waveform, orig_sr=sr, target_sr=target_sr)
    return waveform

def speech_to_text(input_audio_path):
    waveform = resample_audio(input_audio_path)
    input_values = asr_processor(waveform, return_tensors="pt").input_values
    with torch.no_grad():
        logits = asr_model(input_values).logits
    predicted_ids = torch.argmax(logits, dim=-1)
    transcription = asr_processor.batch_decode(predicted_ids)[0]
    return transcription.strip()

def text_to_speech(text, speaker_wav_path, output_path="output_audio/output.wav"):
    if not text.strip():
        return "Empty text input."
    tts.tts_to_file(text=text, file_path=output_path, speaker_wav=speaker_wav_path)
    return output_path

def speech_to_speech(input_audio, text_input=None):
    speaker_wav_path = input_audio
    if text_input is None:
        text_input = speech_to_text(input_audio)
    return text_to_speech(text_input, speaker_wav_path)

iface = gr.Interface(fn=speech_to_speech,
                     inputs=[gr.Audio(type="filepath"), gr.Textbox(optional=True)],
                     outputs=gr.Audio())
iface.launch()