src/text_to_speech.py

import time
import torch
from transformers import set_seed
from transformers import VitsTokenizer, VitsModel

def synthesize_facebook(s:str, iso3:str) -> str:
    ''' 
    For given text, speak it.
    
    Parameters
    ---------- 
    s: str
        The written text.
    is03:str
        The ISO-3 code of the text's language.

    Returns
    ---------- 
    synth:str
        The synthesized audio.
    '''
    
    # Ensure replicability
    set_seed(555) 
    start_time = time.time()
    
    # Load synthesizer
    tokenizer = VitsTokenizer.from_pretrained(f"facebook/mms-tts-{iso3}")
    model = VitsModel.from_pretrained(f"facebook/mms-tts-{iso3}")
    
    inputs = tokenizer(text=s, return_tensors="pt")
    
    # Inference
    with torch.no_grad():
       outputs = model(**inputs)
    
    synth = outputs.waveform[0]    

    print("Time elapsed: ", int(time.time() - start_time), " seconds")
    return synth.numpy()