app.py

import gradio as gr
import numpy as np
import torch
from transformers import pipeline, VitsModel
#from transformers import SpeechT5ForTextToSpeech, SpeechT5HifiGan, SpeechT5Processor
from transformers import WhisperTokenizer, GenerationConfig
from transformers import pipeline, VitsModel, AutoTokenizer, AutoTokenizer

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


tokenizer = WhisperTokenizer.from_pretrained("openai/whisper-medium")
generation_config = GenerationConfig.from_pretrained("openai/whisper-medium")

generation_config.forced_decoder_ids

tokenizer.decode(generation_config.forced_decoder_ids[1][1])


asr_pipe = pipeline("automatic-speech-recognition", model="openai/whisper-medium", device=device)


# ---------------- Speech generator mms-tts-spa --------------------------#

vist_model = VitsModel.from_pretrained("facebook/mms-tts-spa")
vist_tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-spa")

# ---------------- Speech generator  specht5_tts --------------------------#

model = SpeechT5ForTextToSpeech.from_pretrained(
    "juangtzi/speecht5_finetuned_voxpopuli_es"
)
checkpoint = "microsoft/speecht5_tts"
processor = SpeechT5Processor.from_pretrained(checkpoint)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")

speaker_embeddings2 = np.load('speaker_embeddings.npy')
speaker_embeddings2 = torch.tensor(speaker_embeddings2)
print(speaker_embeddings2)


# def language_detector(text):
#     resultado = lang_detector(text)
#     idioma_detectado = resultado[0]['label']
#     print(idioma_detectado)
#     return idioma_detectado

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

# def synthesise(text):
#     inputs = processor(text=text, return_tensors="pt")
#     output = model.generate_speech(inputs["input_ids"], speaker_embeddings2, vocoder=vocoder)
#     return output

# def speech_to_speech_translation(audio):
#     translated_text = translate(audio)
#     synthesised_speech = synthesise(translated_text)
#     audio_data = synthesised_speech.cpu().numpy()
#     #audio_data = np.squeeze(audio_data)
#     #audio_data = audio_data / np.max(np.abs(audio_data))
#     sample_rate = 16000
#     return (sample_rate, audio_data)

def synthesise(text):
    print(text)
    inputs = vist_tokenizer(text, return_tensors="pt")
    with torch.no_grad():
        output = vist_model(**inputs).waveform[0]
    return output

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 Spanish.

![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(sources="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(sources="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()