app.py

import gradio as gr
import subprocess,os
from datasets import load_dataset, Audio
import corpora
import ctcalign,graph
from numpy import random


import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt


def setup():
    r0 = subprocess.run(["pwd"], capture_output=True, text=True)
    print('PWD::', r0.stdout)
    r1 = subprocess.run(["wget", "https://github.com/google/REAPER/archive/refs/heads/master.zip"], capture_output=True, text=True)
    print(r1.stdout)
    subprocess.run(["unzip", "./master.zip"])
    subprocess.run(["mv", "REAPER-master", "REAPER"])
    subprocess.run(["rm", "./master.zip"])
    os.chdir('./REAPER')
    subprocess.run(["mkdir", "build"])
    os.chdir('./build')
    r2 = subprocess.run(["cmake", ".."], capture_output=True, text=True)
    print(r2.stdout)
    r3 = subprocess.run(["make"], capture_output=True, text=True)
    print(r3.stdout)
    
    os.chdir('../..')
    r9 = subprocess.run(["ls", "-la"], capture_output=True, text=True)
    print('LS::', r9.stdout)

                        
#print('about to setup')
setup()

def load_lang(langname):
    if langname=="Icelandic":
        df = corpora.ds_i
        model_path="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-icelandic-ep10-1000h"
    elif langname =="Faroese":
        df = corpora.ds_f
        model_path = "carlosdanielhernandezmena/wav2vec2-large-xlsr-53-faroese-100h"

    model_word_separator = '|'
    model_blank_token = '[PAD]'
    lang_aligner = ctcalign.aligner(model_path,model_word_separator,model_blank_token)

    df = df.data.to_pandas()
    df = df.drop(columns=['audio', 'speaker_id','duration'])
    return (df[:10], lang_aligner) #(df, df[:50])


def f1(langname,lang_aligner):
    if langname=="Icelandic":
        ds = corpora.ds_i
    elif langname =="Faroese":
        ds = corpora.ds_f

    
    #fig = plt.figure(figsize=(10,4))
    #plt.axline((0,0),slope=1,color="darkgray")
    #plt.xlabel("Vowel length (ms)")
    #plt.ylabel("Consonant length (ms)")


    maxdat=len(ds)

    ds = ds.select([random.randint(maxdat-1)])
    #print([th for th in ds.sample()])
    sound_path = ds['audio'][0]['path'] # audio 0 array is the audio data itself
    transcript = ds['normalized_text'][0]
    #print('PLACE A:',lang_aligner)
    return graph.align_and_graph(sound_path,transcript,lang_aligner)


bl = gr.Blocks()

with bl:

    lloadr = gr.Dropdown(["Faroese", "Icelandic"], label="Select a language")#, info="Loading the dataset takes some time")

    align_func = gr.State()#value=ctcalign.aligner(model_path="carlosdanielhernandezmena/wav2vec2-large-xlsr-53-icelandic-ep10-1000h",model_word_separator = '|',model_blank_token = '[PAD]'))
    
    with gr.Row():
        #invisidata = gr.DataFrame(interactive=False, visible=False)
        databrowser = gr.DataFrame(wrap=True, max_rows=50, interactive=False, overflow_row_behaviour='paginate')


    
    btn1 = gr.Button(value="The random prosody button")
    btn1.style(full_width=False, size="sm")

    pl1 = gr.Plot()

    btn1.click(f1, [lloadr,align_func], pl1)



    
    lloadr.change(load_lang,lloadr,[databrowser,align_func])


    gr.Markdown(
        """
    # ABOUT
    This is a work-in-progress demo.
    
    Icelandic uses the [samromur-asr](https://huggingface.co/datasets/language-and-voice-lab/samromur_asr) corpus, and Faroese uses [ravnursson-asr](https://huggingface.co/datasets/carlosdanielhernandezmena/ravnursson_asr).
    
    After you select a language, a few example sentences from the corpus are displayed.
    
    Click the button to view time-aligned prosody information for a random sentence - this could be any sentence, not only one of the ones shown above.
    
    [ABOUT REAPER PITCH TRACKING - TODO]

    [ABOUT RMSE INTENSITY - TODO]

    [ABOUT CTC ALIGNMENT - TODO]

    caitlinr@ru.is / https://github.com/catiR/
    """
    )


if __name__ == "__main__":
    bl.launch()