app.py

import subprocess
import sys
subprocess.check_call([sys.executable, "-m", "pip", "install", 'gradio==3.40.1'])

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="openai/whisper-base", device=device)

# load text-to-speech checkpoint and speaker embeddings
processor = SpeechT5Processor.from_pretrained("tsobolev/speecht5_finetuned_voxpopuli_fi")

model = SpeechT5ForTextToSpeech.from_pretrained("tsobolev/speecht5_finetuned_voxpopuli_fi").to(device)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)

embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
speaker_embeddings = torch.tensor(embeddings_dataset[7000]["xvector"]).unsqueeze(0)
en2fi_pipeline = pipeline("translation", model="Helsinki-NLP/opus-mt-en-fi")


print("gradio version is ",gr.__version__)


def translate(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate"})
    fi_translation = en2fi_pipeline(outputs["text"])
    text = fi_translation[0]['translation_text']
    replacements = [
        ("ä", "ae"),
        ("ö", "oe"),
    ]
    
    for src, dst in replacements:
        text = text.replace(src, dst)

    print(text)
    return 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()


def speech_to_speech_translation(audio):
    translated_text = translate(audio)
    synthesised_speech = synthesise(translated_text)
    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
    return 16000, synthesised_speech


title = "Cascaded STST"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language, supported by the whisper, to target speech in Finnish. 
Demo uses: 
* OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation
* Language Technology at the University of Helsinki en-fi model [Helsinki-NLP](https://huggingface.co/Helsinki-NLP/opus-mt-en-fi) 
* Microsoft's [SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model fine-tuned on subset of [VoxPopuli](https://huggingface.co/datasets/facebook/voxpopuli) dataset for text-to-speech
* replacements: ä => ae , ö => oe
![Cascaded STST](https://huggingface.co/datasets/huggingface-course/audio-course-images/resolve/main/s2st_cascaded.png "Diagram of cascaded speech to speech translation")
"""

demo = gr.Blocks()

mic_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="microphone", type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    title=title,
    description=description,
)

file_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="upload", type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
    examples=[["./example.wav"]],
    title=title,
    description=description,
)

with demo:
    gr.TabbedInterface([mic_translate, file_translate], ["Microphone", "Audio File"])

demo.launch()