testing

app.py

@@ -1,13 +1,99 @@
+import whisper
+import pandas as pd
+import whisper
+import subprocess
+from simple_diarizer.diarizer import Diarizer
 import streamlit as st
 
 
-uploaded_file = st.file_uploader("Choose a file")
 
-number_of_speakers = st.number_input('Insert a number')
+def speech_to_text(uploaded):
+    model = whisper.load_model('tiny')
+    result = model.transcribe(uploaded)
+    return f'You said: {result["text"]}'
 
-st.write('The current number is ', number_of_speakers)
+def segment(nu_speakers):
 
-if uploaded_file is not None:
+    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('base')
+    result = model.transcribe('mono.wav',verbose=True,
+                          without_timestamps=False)
+    tdf = pd.DataFrame(result['segments'])
+    return tdf
+
+def monotize(uploaded):
+    print(uploaded.name)
+    cmd = f"ffmpeg -y -i {uploaded.name} -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:
+        lines.append(row)
+
+    return '\n'.join(lines)
+
+
+descript = ("This web app creates transcripts using OpenAI's [Whisper](https://github.com/openai/whisper) to transcribe "
+            "audio files combined with [Chau](https://github.com/cvqluu)'s [Simple Diarizer](https://github.com/cvqluu/simple_diarizer) "
+            "to partition the text by speaker.\n"
+            "* Creating the transcript takes some time. "
+            "Using the default base transcription model, the process takes approximately 20% of the length of the audio file.\n "
+            "* There seems to be cap on the uploaded file size of about 20MBs. My [colab](https://colab.research.google.com/drive/18AD-mb3bT4s8k3UNhZu-ghPq2DT5il3V?usp=sharing) version "
+            "can handle any file size, but requies some Python knowledge.\n"
+            "* After uploading the file, **be sure to select the number of speakers**." )
+
+
+with st.form(key='my_form'):
+    uploaded = st.file_uploader("Choose a file")
+    nu_speakers = st.slider('Number of speakers in audio file:', min_value=1, max_value=6, value=2, step=1)
+    submit = st.form_submit_button("Transcribe!")
+
+
+if submit:
+    bytes_data = uploaded.getvalue()
+    with open('temp_audio', 'wb') as outfile:
+        outfile.write(bytes_data)
+    speech_to_text('temp_audio')
     # To read file as bytes:
-    bytes_data = uploaded_file.getvalue()
-    st.write(bytes_data)