app.py

import soundfile as sf
import gradio as gr
import jax
import numpy as np
import os
from PIL import Image
import random
import sox
import torch

from transformers import AutoProcessor, AutoModelForCTC
from transformers import pipeline
from dalle_mini import DalleBart, DalleBartProcessor
from vqgan_jax.modeling_flax_vqgan import VQModel

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

api_token = os.getenv("API_TOKEN")
asr_processor = AutoProcessor.from_pretrained("imvladikon/wav2vec2-xls-r-300m-hebrew")
asr_model = AutoModelForCTC.from_pretrained("imvladikon/wav2vec2-xls-r-300m-hebrew")

he_en_translator = pipeline("translation", model="Helsinki-NLP/opus-mt-tc-big-he-en")

# Model references
# dalle-mini, mega too large
# DALLE_MODEL = "dalle-mini/dalle-mini/mega-1-fp16:latest"  # can be wandb artifact or 🤗 Hub or local folder or google bucket
DALLE_MODEL = "dalle-mini/dalle-mini/mini-1:v0"
DALLE_COMMIT_ID = None

# VQGAN model
VQGAN_REPO = "dalle-mini/vqgan_imagenet_f16_16384"
VQGAN_COMMIT_ID = "e93a26e7707683d349bf5d5c41c5b0ef69b677a9"
 
model = DalleBart.from_pretrained(DALLE_MODEL, revision=DALLE_COMMIT_ID)
vqgan = VQModel.from_pretrained(VQGAN_REPO, revision=VQGAN_COMMIT_ID)
processor = DalleBartProcessor.from_pretrained(DALLE_MODEL, revision=DALLE_COMMIT_ID)

def generate_image(text):
    tokenized_prompt = processor([text])

    gen_top_k = None
    gen_top_p = None
    temperature = 0.85
    cond_scale = 3.0
 
    encoded_images = model.generate(
        **tokenized_prompt,
        prng_key=jax.random.PRNGKey(random.randint(0, 1e7)),
        params=model.params,
        top_k=gen_top_k,
        top_p=gen_top_p,
        temperature=temperature,
        condition_scale=cond_scale,
        )
    encoded_images = encoded_images.sequences[..., 1:]
    decoded_images = vqgan.decode_code(encoded_images, vqgan.params)
    decoded_images = decoded_images.clip(0.0, 1.0).reshape((-1, 256, 256, 3))
    img = decoded_images[0]
    return Image.fromarray(np.asarray(img * 255, dtype=np.uint8))

def convert(inputfile, outfile):
    sox_tfm = sox.Transformer()
    sox_tfm.set_output_format(
        file_type="wav", channels=1, encoding="signed-integer", rate=16000, bits=16
    )
    sox_tfm.build(inputfile, outfile)

def parse_transcription(wav_file):
    # Get the wav file from the microphone
    filename = wav_file.name.split('.')[0]
    convert(wav_file.name, filename + "16k.wav")
    speech, _ = sf.read(filename + "16k.wav")

    # transcribe to hebrew
    input_values = asr_processor(speech, sampling_rate=16_000, return_tensors="pt").input_values
    logits = asr_model(input_values).logits
    predicted_ids = torch.argmax(logits, dim=-1)
    transcription = asr_processor.decode(predicted_ids[0], skip_special_tokens=True)
    
    print(transcription)

    # translate to english
    translated = he_en_translator(transcription)[0]['translation_text']

    print(translated) 

    # generate image
    image = generate_image(translated)
    return transcription, image 

outputs = [gr.outputs.Textbox(label="transcript"), gr.outputs.Image(label='')]
input_mic = gr.inputs.Audio(source="microphone", type="file", optional=True)

gr.Interface(parse_transcription, inputs=[input_mic],  outputs=outputs,
             analytics_enabled=False,
             show_tips=False,
             theme='huggingface',
             layout='horizontal',
             title="Draw Me A Sheep in Hebrew",
             enable_queue=True).launch(inline=False)