app.py

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("microsoft/speecht5_tts")
#"ihanif/speecht5_finetuned_voxpopuli_lt"
#model_id = "sanchit-gandhi/speecht5_tts_vox_nl"
model_id = "ihanif/speecht5_finetuned_voxpopuli_lt"
processor = SpeechT5Processor.from_pretrained(model_id)

#model = SpeechT5ForTextToSpeech.from_pretrained("microsoft/speecht5_tts").to(device)
#vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)
model = SpeechT5ForTextToSpeech.from_pretrained(model_id).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[7306]["xvector"]).unsqueeze(0)


replacements = [
    ("à", "a"),
    ("ą", "a"),
    ("ç", "c"),
    ("č", "c"),
    ("è", "e"),
    ("ë", "e"),
    ("ė", "e"),
    ("ę", "e"),
    ("í", "i"),
    ("ï", "i"),
    ("į", "i"),
    ("ö", "o"),
    ("š", "s"),
    ("ü", "u"),
    ("ū", "u"),
    ("ų", "u"),
    ("ž", "z"),
]

def cleanup_text(text):
    for src, dst in replacements:
        text = text.replace(src, dst)
    return text
    
def translate(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "translate", "language": "lt"})
    return outputs["text"]


def synthesise(text):
    text = cleanup_text(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 to target speech in Lithuanian. Demo uses OpenAI's [Whisper Base](https://huggingface.co/openai/whisper-base) model for speech translation, and Microsoft's
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech:

![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"), gr.outputs.Textbox()],
    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"), gr.outputs.Textbox()],
    examples=[["./example.wav"]],
    title=title,
    description=description,
)

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

demo.launch()