app.py

import spaces
import gradio as gr
import numpy as np
import torch

from datasets import load_dataset
from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor, pipeline

device = "cuda:0" if torch.cuda.is_available() else "cpu"


# load speech translation checkpoint
asr_pipe = pipeline("automatic-speech-recognition", model="oyemade/w2v-bert-2.0-yoruba-colab-CV16.1", device=device)


# load text-to-speech checkpoint and speaker embeddings
processor = SpeechT5Processor.from_pretrained("microsoft/speecht5_tts")
model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)

embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation", revision="refs/convert/parquet")
speaker_embeddings = torch.tensor(embeddings_dataset[7306]["xvector"]).unsqueeze(0)


translation_model = pipeline("translation", "facebook/nllb-200-distilled-600M", src_lang="yor_Latn", tgt_lang="eng_Latn", device=device)


def translate(audio):
  text = asr_pipe(audio)["text"]
  # print(text)
  translation = translation_model(text)
  # print(translation[0]['translation_text'])
  return translation[0]['translation_text']

def synthesise(text):
    inputs = processor(text=text, return_tensors="pt")
    speech = model.generate_speech(inputs["input_ids"].to(device), speaker_embeddings.to(device), vocoder=vocoder)
    return speech.cpu()

@spaces.GPU
def speech_to_speech_translation(audio):
  # print(model)
  translated_text = translate(audio)
  synthesised_speech = synthesise(translated_text)
  synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
  return 16000, synthesised_speech
    
iface = gr.Interface(
    speech_to_speech_translation,
    gr.Audio(sources=["microphone", "upload"], type="filepath"),
    gr.Audio(label="Generated Speech", type="numpy"),
    title="Neoform AI: Yoruba Speech to English Speech",
    description="Demo for Yoruba speech translated to English Speech. NOTE: If you get an ERROR after pressing submit, give the audio some secs to load then try again.",
)

iface.launch()