app.py

# -*- coding: utf-8 -*-
"""app.ipynb

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/143eWt9oxUTcF59OBiVybOgKXJB3QOTsK
"""

# Beginning of Unit 7

from transformers.models.markuplm.tokenization_markuplm import MARKUPLM_ENCODE_PLUS_ADDITIONAL_KWARGS_DOCSTRING
import torch, torchaudio
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import sentencepiece
from transformers import MarianMTModel, MarianTokenizer
from datasets import load_dataset
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech, SpeechT5HifiGan
from IPython.display import Audio
import numpy as np

target_dtype = np.int16
max_range = np.iinfo(target_dtype).max

# Load Spanish Audio

def transcribe(audio):
    model_id_asr = "openai/whisper-small"
    processor_asr = WhisperProcessor.from_pretrained(model_id_asr)
    model_asr = WhisperForConditionalGeneration.from_pretrained(model_id_asr)
    model_asr.config.forced_decoder_ids = None

    input_features = processor_asr(audio["audio"]["array"], sampling_rate=audio["audio"]["sampling_rate"], return_tensors="pt").input_features

    predicted_ids = model_asr.generate(input_features)

    # decode token ids to text
    transcription = processor_asr.batch_decode(predicted_ids, skip_special_tokens=True)
    return transcription[0]

# Run inference on Spanish Audio vector

def translate(text):
    model_id_mt = "Helsinki-NLP/opus-mt-es-fr"
    tokenizer_mt = MarianTokenizer.from_pretrained(model_id_mt)
    model_mt = MarianMTModel.from_pretrained(model_id_mt)
    # Tokenize the input text
    input_ids = tokenizer_mt.encode(text, return_tensors="pt")

    # Generate translation
    with torch.no_grad():
        translated_ids = model_mt.generate(input_ids)

    # Decode the translated text
    translated_text = tokenizer_mt.decode(translated_ids[0], skip_special_tokens=True)

    return translated_text


def synthesise(text):

    processor_tts = SpeechT5Processor.from_pretrained("crowbarmassage/speecht5_finetuned_voxpopuli_fr")

    model_tts = SpeechT5ForTextToSpeech.from_pretrained("crowbarmassage/speecht5_finetuned_voxpopuli_fr")
    vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")

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


    # Load your dataset from Hugging Face
    #embeddings_dataset = load_dataset("crowbarmassage/MAEmbed")
    #print(embeddings_dataset.features)
    #print(embeddings_dataset[0])

    # Extract the embedding (assuming it's in a column named 'embedding')
    # Note: Adjust the index [0] if your embedding is at a different position in the dataset.
    #embedding_array = embeddings_dataset[0]['embedding']

    # Convert the embedding to a PyTorch tensor and add a batch dimension
    #speaker_embeddings = torch.tensor(embedding_array).unsqueeze(0)
    print(speaker_embeddings)
    print(type(speaker_embeddings))
    inputs = processor_tts(text=text, return_tensors="pt")
    speech = model_tts.generate_speech(
        inputs["input_ids"], speaker_embeddings, vocoder=vocoder
    )
    print(speech)
    print(type(speech))
    return speech

def speech_to_speech_translation(audio_filepath):
    # Load the audio file
    waveform, sampling_rate = torchaudio.load(audio_filepath)

    if sampling_rate != 16000:
      resampler = torchaudio.transforms.Resample(orig_freq=sampling_rate, new_freq=16000)
      waveform = resampler(waveform)
      sampling_rate = 16000
    # Convert the waveform to a numpy array and construct the expected dictionary format
    audio_dict = {
        "audio": {
            "array": waveform.numpy(),
            "sampling_rate": sampling_rate
        }
    }
    transcribed_text = transcribe(audio_dict)
    translated_text = translate(transcribed_text)
    synthesised_speech = synthesise(translated_text)
    #print(transcribed_text)
    #print(translated_text)
    #print(synthesised_speech)
    #print(torch.min(synthesised_speech), torch.max(synthesised_speech))
    synthesised_speech = (synthesised_speech * 32767).numpy().astype(np.int16)
    #print(synthesised_speech)
    #print(np.min(synthesised_speech), np.max(synthesised_speech))
    return 16000, synthesised_speech

import gradio as gr

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"),
)

file_translate = gr.Interface(
    fn=speech_to_speech_translation,
    inputs=gr.Audio(source="upload", type="filepath"),
    outputs=gr.Audio(label="Generated Speech", type="numpy"),
)

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

demo.launch(debug=True, share=False)