app.py

import gradio as gr
from transformers import Wav2Vec2Processor
from transformers import AutoModelForCTC
from conversationalnlp.models.wav2vec2 import Wav2Vec2Predict
from conversationalnlp.models.wav2vec2 import ModelLoader
from conversationalnlp.utils import *
import soundfile as sf
import os

"""
run gradio with 
>>python app.py
"""

audiosavepath = os.getcwd()

pretrained_model = "codenamewei/speech-to-text"

processor = Wav2Vec2Processor.from_pretrained(
    pretrained_model)

model = AutoModelForCTC.from_pretrained(
    pretrained_model)

modelloader = ModelLoader(model, processor)

predictor = Wav2Vec2Predict(modelloader)


def greet(audioarray):
    """
    audio array in the following format

    (16000, array([ -5277184,    326400,   -120320, ...,  -5970432, -12745216,
        -6934528], dtype=int32))
    <class 'tuple'>
    """
    audioabspath = os.path.join(audiosavepath, "temp.wav")

    # WORKAROUND: Save to file and reread to get the array shape needed for prediction
    sf.write(audioabspath, audioarray[1], audioarray[0])

    print(f"Audio at path {audioabspath}")
    predictiontexts = predictor.predictfiles([audioabspath])
    outputtext = predictiontexts["predicted_text"][-1] + \
        "\n" + predictiontexts["corrected_text"][-1]

    return outputtext


demo = gr.Interface(fn=greet, inputs="audio",
                    outputs="text", title="Speech-to-Text")

demo.launch()  # share=True)