speech-to-chat / app.py
kobakhit's picture
gpt version
af8cb29
import streamlit as st
import streamlit_ext as ste
import openai
from pydub import AudioSegment
# from pytube import YouTube
# import pytube
import yt_dlp
import io
from pyannote.audio import Pipeline
from pyannote.audio.pipelines.utils.hook import ProgressHook
from pyannote.database.util import load_rttm
from pyannote.core import Annotation, Segment, notebook
import time
import json
import torch
import urllib.parse as urlparse
from urllib.parse import urlencode
import os
import unicodedata
import re
import matplotlib
matplotlib.use('Agg')
from matplotlib import pyplot as plt
st.set_page_config(
page_title="Speech-to-chat",
page_icon = '🌊',
layout='wide'
)
# Set your OpenAI, Hugging Face API keys
try:
openai.api_key = st.secrets['openai']
hf_api_key = st.secrets['hf']
except Exception:
openai.api_key = os.getenv['openai']
hf_api_key = os.getenv['hf']
TRANSCRIPTION_REQUEST_LIMIT = 550
PROMPT_REQUEST_LIMIT = 20
DURATION_LIMIT = 3600 # seconds
def create_audio_stream(audio):
return io.BytesIO(audio.export(format="wav").read())
def add_query_parameter(link, params):
url_parts = list(urlparse.urlparse(link))
query = dict(urlparse.parse_qsl(url_parts[4]))
query.update(params)
url_parts[4] = urlencode(query)
return urlparse.urlunparse(url_parts)
def slugify(value, allow_unicode=False):
"""
Taken from https://github.com/django/django/blob/master/django/utils/text.py
Convert to ASCII if 'allow_unicode' is False. Convert spaces or repeated
dashes to single dashes. Remove characters that aren't alphanumerics,
underscores, or hyphens. Convert to lowercase. Also strip leading and
trailing whitespace, dashes, and underscores.
"""
value = str(value)
if allow_unicode:
value = unicodedata.normalize('NFKC', value)
else:
value = unicodedata.normalize('NFKD', value).encode('ascii', 'ignore').decode('ascii')
value = re.sub(r'[^\w\s-]', '', value.lower())
return re.sub(r'[-\s]+', '-', value).strip('-_')
def youtube_video_id(value):
"""
Examples:
- http://youtu.be/SA2iWivDJiE
- http://www.youtube.com/watch?v=_oPAwA_Udwc&feature=feedu
- http://www.youtube.com/embed/SA2iWivDJiE
- http://www.youtube.com/v/SA2iWivDJiE?version=3&hl=en_US
"""
query = urlparse.urlparse(value)
if query.hostname == 'youtu.be':
return query.path[1:]
if query.hostname in ('www.youtube.com', 'youtube.com'):
if query.path == '/watch':
p = urlparse.parse_qs(query.query)
return p['v'][0]
if query.path[:7] == '/embed/':
return query.path.split('/')[2]
if query.path[:3] == '/v/':
return query.path.split('/')[2]
# fail?
return None
@st.cache_data
def process_youtube_link2(youtube_link):
'''
uses pytube https://github.com/pytube/pytube
issue with https://github.com/pytube/pytube/issues/84
'''
try:
yt = YouTube(youtube_link)
audio_stream = yt.streams.filter(only_audio=True).first()
audio_name = audio_stream.default_filename
st.write(f"Downloaded {audio_name}")
except pytube.exceptions.AgeRestrictedError:
st.warning('Age restricted videos cannot be processed.')
st.stop()
try:
os.remove('sample.mp4')
except OSError:
pass
audio_file = audio_stream.download(filename='sample.mp4')
time.sleep(2)
audio = load_audio('sample.mp4')
st.audio(create_audio_stream(audio), format="audio/mp4", start_time=0)
return audio, audio_name
@st.cache_data
def process_youtube_link(youtube_link):
'uses yt-dlp https://github.com/yt-dlp/yt-dlp'
try:
os.remove('sample.m4a')
except OSError:
pass
ydl_opts = {
'format': 'm4a/bestaudio/best',
# ℹ️ See help(yt_dlp.postprocessor) for a list of available Postprocessors and their arguments
'outtmpl': './sample.%(ext)s'
# 'postprocessors': [{ # Extract audio using ffmpeg
# 'key': 'FFmpegExtractAudio',
# 'preferredcodec': 'm4a',
# }]
}
try:
with yt_dlp.YoutubeDL(ydl_opts) as ydl:
info = ydl.extract_info(youtube_link, download=True)
audio_name = slugify( info['title'] )
st.write(f"Downloaded {info['title']}")
except Exception as e:
st.warning(e)
st.stop()
audio = load_audio(f'sample.m4a')
st.audio(create_audio_stream(audio), format="audio/m4a", start_time=0)
return audio, audio_name
@st.cache_data
def load_rttm_file(rttm_path):
return load_rttm(rttm_path)['stream']
def load_audio(uploaded_audio):
return AudioSegment.from_file(uploaded_audio)
if "openai_model" not in st.session_state:
st.session_state["openai_model"] = "gpt-4o-mini"
if "prompt_request_counter" not in st.session_state:
st.session_state["prompt_request_counter"] = 0
initial_prompt = [{"role": "system", "content": "You are helping to analyze and summarize a transcript of a conversation."},
{"role": 'user', "content": 'Please summarize briefly below transcript and inlcude a list of tags with a hash for SEO. \n{}'}]
if "messages" not in st.session_state:
st.session_state.messages = initial_prompt
st.title("Speech-to-Chat")
reddit_thread = 'https://www.reddit.com/r/dataisbeautiful/comments/17413bq/oc_speech_diarization_app_that_transcribes_audio'
with st.sidebar:
st.markdown('''
# How to Use
1. Enter a youtube link.
2. "Chat" with the video.
Example prompts:
- Which speaker spoke the most?
- Give me a list of tags with a hash for SEO based on this transcript.
''')
api_key_input = st.text_input(
"OpenAI API Key to lift request limits (Coming soon)",
disabled=True,
type="password",
placeholder="Paste your OpenAI API key here (sk-...)",
help="You can get your API key from https://platform.openai.com/account/api-keys.", # noqa: E501
value=os.environ.get("OPENAI_API_KEY", None)
or st.session_state.get("OPENAI_API_KEY", ""),
)
st.divider()
st.markdown(f'''
# About
Given an audio file or a youtube link this app will
- [x] 1. Partition the audio according to the identity of each speaker (diarization) using `pyannote` [HuggingFace Speaker Diarization api](https://huggingface.co/pyannote/speaker-diarization-3.0)
- [x] 2. Transcribe each audio segment using [OpenAi Whisper API](https://platform.openai.com/docs/guides/speech-to-text/quickstart)
- [x] 3. Set up an LLM chat with the transcript loaded into its knowledge database, so that a user can "talk" to the transcript of the audio file.
This version will only process up to first 6 minutes of an audio file due to limited resources of free tier Streamlit.io/HuggingFace Spaces.
A local version with access to a GPU can process 1 hour of audio in 1 to 5 minutes.
If you would like to use this app at scale reach out directly by creating an issue on [github🤖](https://github.com/KobaKhit/speech-to-text-app/issues)!
Rule of thumb, for this free tier hosted app it takes half the duration of the audio to complete processing, ex. g. 6 minute youtube video will take 3 minutes to diarize.
Made by [kobakhit](https://github.com/KobaKhit/speech-to-text-app)
''')
# Chat container
container_transcript_chat = st.container()
# Source Selection
option = st.radio("Select source:", [ "Use YouTube link","See Example"], index=0)
# Upload audio file
if option == "Upload an audio file":
with st.form('uploaded-file', clear_on_submit=True):
uploaded_audio = st.file_uploader("Upload an audio file (MP3 or WAV)", type=["mp3", "wav","mp4"])
st.form_submit_button()
if st.form_submit_button(): st.session_state.messages = initial_prompt
with st.expander('Optional Parameters'):
# st.session_state.rttm = st.file_uploader("Upload .rttm if you already have one", type=["rttm"])
# st.session_state.transcript_file = st.file_uploader("Upload transcipt json", type=["json"])
youtube_link = st.text_input('Youtube link of the audio sample')
if uploaded_audio is not None:
st.audio(uploaded_audio, format="audio/wav", start_time=0)
audio_name = uploaded_audio.name
audio = load_audio(uploaded_audio)
# sample_rate = st.number_input("Enter the sample rate of the audio", min_value=8000, max_value=48000)
# audio = audio.set_frame_rate(sample_rate)
# use youtube link
elif option == "Use YouTube link":
with st.form('youtube-link'):
youtube_link_raw = st.text_input("Enter the YouTube video URL:")
youtube_link = f'https://youtu.be/{youtube_video_id(youtube_link_raw)}'
if st.form_submit_button(): # reset variables on new link submit
process_youtube_link.clear()
st.session_state.messages = initial_prompt
st.session_state.rttm = None
st.session_state.transcript_file = None
st.session_state.prompt_request_counter = 0
with container_transcript_chat:
st.empty()
# with st.expander('Optional Parameters'):
# st.session_state.rttm = st.file_uploader("Upload .rttm if you already have one", type=["rttm"])
# st.session_state.transcript_file = st.file_uploader("Upload transcipt json", type=["json"])
if youtube_link_raw:
audio, audio_name = process_youtube_link(youtube_link)
# sample_rate = st.number_input("Enter the sample rate of the audio", min_value=8000, max_value=48000)
# audio = audio.set_frame_rate(sample_rate)
# except Exception as e:
# st.write(f"Error: {str(e)}")
elif option == 'See Example':
youtube_link = 'https://www.youtube.com/watch?v=TamrOZX9bu8'
audio_name = 'Stephen A. Smith has JOKES with Shannon Sharpe'
st.write(f'Loaded audio file from {youtube_link} - {audio_name} 👏😂')
if os.path.isfile('example/steve a smith jokes.mp4'):
audio = load_audio('example/steve a smith jokes.mp4')
else:
yt = YouTube(youtube_link)
audio_stream = yt.streams.filter(only_audio=True).first()
audio_file = audio_stream.download(filename='sample.mp4')
time.sleep(2)
audio = load_audio('sample.mp4')
if os.path.isfile("example/steve a smith jokes.rttm"):
st.session_state.rttm = "example/steve a smith jokes.rttm"
if os.path.isfile('example/steve a smith jokes.json'):
st.session_state.transcript_file = 'example/steve a smith jokes.json'
st.audio(create_audio_stream(audio), format="audio/mp4", start_time=0)
# Diarize
if "audio" in locals():
# create stream
duration = audio.duration_seconds
if duration > DURATION_LIMIT:
st.info(f'Only processing the first {int(DURATION_LIMIT/6/6)} minutes of the audio due to Streamlit.io resource limits.')
audio = audio[:DURATION_LIMIT*1000]
duration = audio.duration_seconds
# Perform diarization with PyAnnote
pipeline = Pipeline.from_pretrained(
"pyannote/speaker-diarization-3.0", use_auth_token=hf_api_key)
if torch.cuda.device_count() > 0: # use gpu if available
st.write('Using cuda - GPU')
pipeline.to(torch.device('cuda'))
# run the pipeline on an audio file
with st.spinner('Performing Diarization...'):
if 'rttm' in st.session_state and st.session_state.rttm != None:
st.write(f'Loading {st.session_state.rttm}')
diarization = load_rttm_file(st.session_state.rttm )
else:
# make progress hook
# with ProgressHook() as hook:
# diarization = pipeline(audio_, hook=hook)
diarization = pipeline(create_audio_stream(audio))
# dump the diarization output to disk using RTTM format
with open(f'{audio_name.split(".")[0]}.rttm', "w") as f:
diarization.write_rttm(f)
st.session_state.rttm = f'{audio_name.split(".")[0]}.rttm'
# Display the diarization results
st.write("Diarization Results:")
annotation = Annotation()
sp_chunks = []
progress_text = f"Processing 1/{len(sp_chunks)}..."
my_bar = st.progress(0, text=progress_text)
counter = 0
n_tracks = len([a for a in diarization.itertracks(yield_label=True)])
for turn, _, speaker in diarization.itertracks(yield_label=True):
annotation[turn] = speaker
progress_text = f"Processing {counter}/{len(sp_chunks)}..."
my_bar.progress((counter+1)/n_tracks, text=progress_text)
counter +=1
temp = {'speaker': speaker,
'start': turn.start, 'end': turn.end, 'duration': turn.end-turn.start,
'audio': audio[turn.start*1000:turn.end*1000]}
if 'transcript_file' in st.session_state and st.session_state.transcript_file == None:
temp['audio_stream'] = create_audio_stream(audio[turn.start*1000:turn.end*1000])
sp_chunks.append(temp)
# plot
notebook.crop = Segment(-1, duration + 1)
figure, ax = plt.subplots(figsize=(10,3))
notebook.plot_annotation(annotation, ax=ax, time=True, legend=True)
figure.tight_layout()
# save to file
st.pyplot(figure)
st.write('Speakers and Audio Samples')
with st.expander('Samples', expanded=True):
for speaker in set(s['speaker'] for s in sp_chunks):
temp = max(filter(lambda d: d['speaker'] == speaker, sp_chunks), key=lambda x: x['duration'])
speak_time = sum(c['duration'] for c in filter(lambda d: d['speaker'] == speaker, sp_chunks))
rate = 100*min((speak_time, duration))/duration
speaker_summary = f"{temp['speaker']} ({round(rate)}% of video duration): start={temp['start']:.1f}s stop={temp['end']:.1f}s"
if youtube_link != None:
speaker_summary += f" {add_query_parameter(youtube_link, {'t':str(int(temp['start']))})}"
st.write(speaker_summary)
st.audio(create_audio_stream(temp['audio']))
st.divider()
# # Perform transcription with Whisper ASR
# Transcript containers
st.write(f'Transcribing using Whisper API ({TRANSCRIPTION_REQUEST_LIMIT} requests limit)...')
container_transcript_completed = st.container()
progress_text = f"Processing 1/{len(sp_chunks[:TRANSCRIPTION_REQUEST_LIMIT])}..."
my_bar = st.progress(0, text=progress_text)
# rework the loop. Simplify if Else
with st.expander('Transcript', expanded=True):
if 'transcript_file' in st.session_state and st.session_state.transcript_file != None:
with open(st.session_state.transcript_file,'r') as f:
sp_chunks_loaded = json.load(f)
for i,s in enumerate(sp_chunks_loaded):
if s['transcript'] != None:
transcript_summary = f"**{s['speaker']}** start={float(s['start']):.1f}s end={float(s['end']):.1f}s: {s['transcript']}"
if youtube_link != None and youtube_link != '':
transcript_summary += f" {add_query_parameter(youtube_link, {'t':str(int(s['start']))})}"
st.markdown(transcript_summary)
progress_text = f"Processing {i+1}/{len(sp_chunks_loaded)}..."
my_bar.progress((i+1)/len(sp_chunks_loaded), text=progress_text)
transcript_json = sp_chunks_loaded
transcript_path = f'{audio_name.split(".")[0]}-transcript.json'
else:
sp_chunks_updated = []
for i,s in enumerate(sp_chunks[:TRANSCRIPTION_REQUEST_LIMIT]):
if s['duration'] > 0.1:
audio_path = s['audio'].export('temp.wav',format='wav')
try:
transcript = openai.Audio.transcribe("whisper-1", audio_path)['text']
except Exception:
transcript = ''
pass
if transcript !='' and transcript != None:
s['transcript'] = transcript
transcript_summary = f"**{s['speaker']}** start={s['start']:.1f}s end={s['end']:.1f}s : {s['transcript']}"
if youtube_link != None:
transcript_summary += f" {add_query_parameter(youtube_link, {'t':str(int(s['start']))})}"
sp_chunks_updated.append({'speaker':s['speaker'],
'start':s['start'], 'end':s['end'],
'duration': s['duration'],'transcript': transcript})
st.markdown(transcript_summary)
progress_text = f"Processing {i+1}/{len(sp_chunks[:TRANSCRIPTION_REQUEST_LIMIT])}..."
my_bar.progress((i+1)/len(sp_chunks[:TRANSCRIPTION_REQUEST_LIMIT]), text=progress_text)
transcript_json = [dict((k, d[k]) for k in ['speaker','start','end','duration','transcript'] if k in d) for d in sp_chunks_updated]
transcript_path = f'{audio_name.split(".")[0]}-transcript.json'
st.session_state.transcript_file = transcript_path
# save the trancript file
with open(transcript_path,'w') as f:
json.dump(transcript_json, f)
# generate transcript string
transcript_string = '\n'.join([f"{s['speaker']} start={s['start']:.1f}s end={s['end']:.1f}s : {s['transcript']}" for s in transcript_json])
@st.cache_data
def get_initial_response(transcript_string):
st.session_state.messages[1]['content'] = st.session_state.messages[1]['content'].format(transcript_string)
initial_response = openai.ChatCompletion.create(
model=st.session_state["openai_model"],
messages=st.session_state.messages
)
return initial_response['choices'][0]['message']['content']
# Chat container
st.session_state.messages[1]['content'] = st.session_state.messages[1]['content'].format(transcript_string)
with container_transcript_chat:
# get a summary of transcript from ChatGpt
try:
init = get_initial_response(transcript_string)
except openai.error.APIError:
# st.stop('It is not you. It is not this app. It is OpenAI API thats having issues.')
init = ''
st.warning('OpenAI API is having issues. Hope they resolve it soon. Refer to https://status.openai.com/')
# pass transcript to initial prompt
# LLM Chat
with st.expander('Summary of the Transcribed Audio File Generated by [`gpt-40-mini`](https://platform.openai.com/docs/models/gpt-4o-mini)', expanded = True):
# display the AI generated summary.
with st.chat_message("assistant", avatar='https://upload.wikimedia.org/wikipedia/commons/0/04/ChatGPT_logo.svg'):
st.write(init)
# chat field
with st.form("Chat",clear_on_submit=True):
prompt = st.text_input(f'Chat with the Transcript ({int(PROMPT_REQUEST_LIMIT)} prompts limit)')
st.form_submit_button()
# message list
# for message in st.session_state.messages[2:]:
# with st.chat_message(message["role"]):
# st.markdown(message["content"])
# make request if prompt was entered
if prompt:
st.session_state.prompt_request_counter += 1
if st.session_state.prompt_request_counter > PROMPT_REQUEST_LIMIT:
st.warning('Exceeded prompt limit.');
st.stop()
# append user prompt to messages
st.session_state.messages.append({"role": "user", "content": prompt})
# dislay user prompt
with st.chat_message("user"):
st.markdown(prompt)
# stream LLM Assisstant response
with st.chat_message("assistant"):
message_placeholder = st.empty()
full_response = ""
# stream response
for response in openai.ChatCompletion.create(
model=st.session_state["openai_model"],
messages=[
{"role": m["role"], "content": m["content"]}
for m in st.session_state.messages
],
stream=True,
):
full_response += response.choices[0].delta.get("content", "")
message_placeholder.markdown(full_response + "▌")
message_placeholder.markdown(full_response)
# append ai response to messages
st.session_state.messages.append({"role": "assistant", "content": full_response})
# Trancription Completed Section
with container_transcript_completed:
st.info(f'Completed transcribing')
@st.cache_data
def convert_df(string):
# IMPORTANT: Cache the conversion to prevent computation on every rerun
return string.encode('utf-8')
# encode transcript string
transcript_json_download = convert_df(json.dumps(transcript_json))
# transcript download buttons
c1_b,c2_b = st.columns((1,1))
# json button
with c1_b:
ste.download_button(
"Download transcript as json",
transcript_json_download,
transcript_path,
)
# create csv string
header = ','.join(transcript_json[0].keys()) + '\n'
for s in transcript_json:
header += ','.join([str(e) if ',' not in str(e) else '"' + str(e) + '"' for e in s.values()]) + '\n'
# csv button
transcript_csv_download = convert_df(header)
with c2_b:
ste.download_button(
"Download transcript as csv",
transcript_csv_download,
f'{audio_name.split(".")[0]}-transcript.csv'
)