app.py

import whisper
import pandas as pd
import whisper
import subprocess
from simple_diarizer.diarizer import Diarizer
import streamlit as st

def speech_to_text(uploaded):
    model = whisper.load_model('tiny')
    result = model.transcribe(uploaded,verbose=True)
    return f'You said: {result["text"]}'

def segment(nu_speakers):

    diar = Diarizer(embed_model='xvec',cluster_method='sc')
    segments = diar.diarize('mono.wav', num_speakers=nu_speakers)

    sdf = pd.DataFrame(segments)

    # reorganize so the first speaker is always speaker 1
    speaker_s = sdf['label'].drop_duplicates().reset_index()['label']
    speaker_d = dict((v,k+1) for k,v in speaker_s.items())

    sdf['speaker'] = sdf['label'].replace(speaker_d)
    return sdf

def audio_to_df(uploaded):
    monotize(uploaded)
    model = whisper.load_model('tiny')
    result = model.transcribe('mono.wav',verbose=True,
                          without_timestamps=False)
    tdf = pd.DataFrame(result['segments'])
    return tdf

def monotize(uploaded):
    cmd = f"ffmpeg -y -i {uploaded} -acodec pcm_s16le -ar 16000 -ac 1 mono.wav"
    subprocess.Popen(cmd, shell=True).wait()

def add_preface(row):
    text = row['text'].replace('\n','')
    speaker = row['speaker']
    return f'Speaker {speaker}: {text}'

def transcribe(uploaded, nu_speakers):

    monotize(uploaded)
    tdf = audio_to_df(uploaded)
    sdf = segment(nu_speakers)

    ns_list = sdf[['start','speaker']].to_dict(orient='records')

    # Find the nearest transcript line to the start of each speaker
    for row in ns_list:
        input = row['start']
        id = tdf.iloc[(tdf['start']-input).abs().argsort()[:1]]['id'].values[0]
        tdf.loc[tdf['id'] ==id, 'speaker'] = row['speaker']

    tdf['speaker'].fillna(method = 'ffill', inplace = True)
    tdf['speaker'].fillna(method = 'bfill', inplace = True)

    tdf['n1'] = tdf['speaker'] != tdf['speaker'].shift(1)
    tdf['speach'] = tdf['n1'].cumsum()
    binned_df = tdf.groupby(['speach', 'speaker'])['text'].apply('\n'.join).reset_index()

    binned_df['speaker'] = binned_df['speaker'].astype(int)
    binned_df['output'] = binned_df.apply(add_preface, axis=1)

    lines = []
    for row in binned_df['output'].values:
        st.write(row)
        lines.append(row)

    return '\n'.join(lines)


form = st.form(key='my_form')
uploaded = form.file_uploader("Choose a file")
nu_speakers = form.slider('Number of speakers in audio file:', min_value=1, max_value=6, value=2, step=1)
submit = form.form_submit_button("Transcribe!")


if submit:
    bytes_data = uploaded.getvalue()
    with open('temp_audio', 'wb') as outfile:
        outfile.write(bytes_data)
    st.write('Converting audio file.')
    monotize('temp_audio')
    text = transcribe('temp_audio', nu_speakers)