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-large-v2", device=device)

# load text-to-speech checkpoint and speaker embeddings
model_id = "Sandiago21/speecht5_finetuned_mozilla_foundation_common_voice_13_german"  # update with your model id
# pipe = pipeline("automatic-speech-recognition", model=model_id)
model = SpeechT5ForTextToSpeech.from_pretrained(model_id)
processor = SpeechT5Processor.from_pretrained(model_id)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
speaker_embeddings = torch.tensor(embeddings_dataset[7440]["xvector"]).unsqueeze(0)

replacements = [    
    ("Ä", "E"),
    ("Æ", "E"),
    ("Ç", "C"),
    ("É", "E"),
    ("Í", "I"),
    ("Ó", "O"),
    ("Ö", "E"),
    ("Ü", "Y"),
    ("ß", "S"),
    ("à", "a"),
    ("á", "a"),
    ("ã", "a"),
    ("ä", "e"),
    ("å", "a"),
    ("ë", "e"),
    ("í", "i"),
    ("ï", "i"),
    ("ð", "o"),
    ("ñ", "n"),
    ("ò", "o"),
    ("ó", "o"),
    ("ô", "o"),
    ("ö", "u"),
    ("ú", "u"),
    ("ü", "y"),
    ("ý", "y"),
    ("Ā", "A"),
    ("ā", "a"),
    ("ă", "a"),
    ("ą", "a"),
    ("ć", "c"),
    ("Č", "C"),
    ("č", "c"),
    ("ď", "d"),
    ("Đ", "D"),
    ("ę", "e"),
    ("ě", "e"),
    ("ğ", "g"),
    ("İ", "I"),
    ("О", "O"),
    ("Ł", "L"),
    ("ń", "n"),
    ("ň", "n"),
    ("Ō", "O"),
    ("ō", "o"),
    ("ő", "o"),
    ("ř", "r"),
    ("Ś", "S"),
    ("ś", "s"),
    ("Ş", "S"),
    ("ş", "s"),
    ("Š", "S"),
    ("š", "s"),
    ("ū", "u"),
    ("ź", "z"),
    ("Ż", "Z"),
    ("Ž", "Z"),
    ("ǐ", "i"),
    ("ǐ", "i"),
    ("ș", "s"),
    ("ț", "t"),
]


def cleanup_text(text):
    for src, dst in replacements:
        text = text.replace(src, dst)
    return text


def transcribe_to_german(audio):
    outputs = asr_pipe(audio, max_new_tokens=256, generate_kwargs={"task": "transcribe", "language": "german"})
    return outputs["text"]


def synthesise_from_german(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 = transcribe_to_german(audio)
    synthesised_speech = synthesise_from_german(translated_text)
    synthesised_speech = (synthesised_speech.numpy() * 32767).astype(np.int16)
    return ((16000, synthesised_speech), translated_text)


title = "Cascaded STST"
description = """
Demo for cascaded speech-to-speech translation (STST), mapping from source speech in any language to target speech in German. Demo uses OpenAI's [Whisper Large v2](https://huggingface.co/openai/whisper-large-v2) model for speech translation, and [Sandiago21/speecht5_finetuned_mozilla_foundation_common_voice_13_german](https://huggingface.co/Sandiago21/speecht5_finetuned_mozilla_foundation_common_voice_13_german) checkpoint for text-to-speech, which is based on Microsoft's
[SpeechT5 TTS](https://huggingface.co/microsoft/speecht5_tts) model for text-to-speech, fine-tuned in German Audio dataset:
![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()