我正在开发一个处理 MP3 文件的 Streamlit 应用程序。主要步骤包括：
上传 MP3 文件。 使用 pydub 将音频分割成更小的块。 使用 OpenAI 转录这些块。 总结使用 Transformer 和 OpenAI 的转录。 --> 这就是我遇到麻烦的地方！！
虽然我已经建立了一个尝试使用 Streamlit 的 session_state 按顺序处理每个阶段的系统，但我遇到了意外的行为。
转录后，不触发摘要。应用程序刚刚刷新，我就收到消息：“发生错误：名称‘summarized_text’未定义”（上传小于 25MB 的小型 mp3 文件时会发生这种情况）
但是当上传大型 mp3 文件（>25MB）时，我立即收到消息：“发生错误：名称‘转录’未定义”。
我想我以某种方式弄乱了状态和分块部分......
有人有想法或可以指出我可能出错的地方吗？任何指导将不胜感激！
import streamlit as st
from pydub import AudioSegment
from pydub.silence import split_on_silence
import os
import openai
from transformers import T5Tokenizer, T5ForConditionalGeneration, GPT2TokenizerFast, pipeline
import textwrap
from concurrent.futures import ThreadPoolExecutor
import warnings
warnings.filterwarnings("ignore")
# Get the password from Streamlit secrets
correct_password = st.secrets["password"]["value"]
password_placeholder = st.empty()
password = password_placeholder.text_input("Enter the password", type="password")
if password != correct_password:
st.error("The password you entered is incorrect.")
st.stop()
# Get the OpenAI key from Streamlit secrets
openai.api_key = st.secrets["openai"]["key"]
def split_audio(file_path, min_silence_len=500, silence_thresh=-40, chunk_length=30000):
st.write("Splitting audio into smaller chunks...")
progress_bar = st.progress(0)
# Load audio file
audio = AudioSegment.from_mp3(file_path)
# Split audio into chunks based on silence
chunks = split_on_silence(
audio,
min_silence_len=min_silence_len,
silence_thresh=silence_thresh,
keep_silence=100
)
# If chunks are longer than desired chunk_length, split them further
split_chunks = []
for i, chunk in enumerate(chunks):
if len(chunk) > chunk_length:
num_mini_chunks = len(chunk) // chunk_length
for j in range(num_mini_chunks):
start_time = j * chunk_length
end_time = start_time + chunk_length
split_chunks.append(chunk[start_time:end_time])
else:
split_chunks.append(chunk)
# Here's where you'd update the progress bar within the function
progress_bar.progress(i / len(chunks))
return split_chunks
st.title("Transcription and Summary App")
# Initialize the processing stage if it's not set
if "stage" not in st.session_state:
st.session_state.stage = 0
# Stage 0: Wait for the user to upload a file
if st.session_state.stage == 0:
audio_file = st.file_uploader("Upload MP3 Audio File", type=["mp3"])
if audio_file is not None:
st.session_state.stage = 1
# Stage 1: Transcribe the audio
if st.session_state.stage == 1:
if audio_file is not None:
try:
# Write to a temp file
with open("temp.mp3", "wb") as f:
f.write(audio_file.getbuffer())
# Splitting the audio into smaller chunks if file size exceeds 25MB
audio_file_size = os.path.getsize("temp.mp3")
if audio_file_size > 25 * 1024 * 1024: # 25MB in bytes
if st.button("start transcription now"):
progress_bar = st.progress(0)
chunks = split_audio("temp.mp3")
progress_bar = st.progress(0)
transcriptions = []
for i, chunk in enumerate(chunks):
progress_bar.progress(i / len(chunks))
with open("temp_chunk.mp3", "wb") as f:
chunk.export(f, format="mp3")
with open("temp_chunk.mp3", "rb") as audio:
transcription_chunk = openai.Audio.translate("whisper-1", audio)["text"]
transcriptions.append(transcription_chunk)
transcription = " ".join(transcriptions)
else:
with open("temp.mp3", "rb") as audio:
transcription = openai.Audio.translate("whisper-1", audio)["text"]
st.write("Transcription: ", transcription)
st.session_state.stage = 2
except Exception as e:
st.write("An error occurred: ", str(e))
# Stage 2: Summarize the transcription
if st.session_state.stage == 2:
try:
if st.button("summarize now"):
st.write("Summarized Text: ", summarized_text)
st.session_state.stage = 3 # or reset to 0 if you want the process to be repeatable
except Exception as e:
st.write("An error occurred: ", str(e))
# Functions for token count, truncation, summarization, etc.
def count_tokens(input_data, max_tokens=20000, input_type='text'):
tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
if input_type == 'text':
tokens = tokenizer.tokenize(input_data)
elif input_type == 'tokens':
tokens = input_data
else:
raise ValueError("Invalid input_type. Must be 'text' or 'tokens'")
token_count = len(tokens)
return token_count
def truncate_text_by_tokens(text, max_tokens=3000):
tokenizer = GPT2TokenizerFast.from_pretrained("gpt2")
tokens = tokenizer.tokenize(text)
truncated_tokens = tokens[:max_tokens]
trunc_token_len = count_tokens(truncated_tokens, input_type='tokens')
truncated_text = tokenizer.convert_tokens_to_string(truncated_tokens)
return truncated_text
def summarize_chunk(classifier, chunk):
summary = classifier(chunk)
return summary[0]["summary_text"]
def summarize_text(text, model_name="t5-small", max_workers=8):
classifier = pipeline("summarization", model=model_name)
summarized_text = ""
chunks = textwrap.wrap(text, width=500, break_long_words=False)
with ThreadPoolExecutor(max_workers=max_workers) as executor:
summaries = executor.map(lambda chunk: summarize_chunk(classifier, chunk), chunks)
summarized_text = " ".join(summaries)
text_len_in_tokens = count_tokens(text)
summary_token_len = count_tokens(summarized_text)
if summary_token_len > 2500:
summarized_text = truncate_text_by_tokens(summarized_text, max_tokens=2500)
with open("transcript_summary.txt", "w") as file:
file.write(summarized_text)
return summarized_text.strip()
def gpt_summarize_transcript(transcript_text, token_len):
response = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=[
{"role": "system", "content": "You are an expert at summarizing long documents into concise and comprehensive summaries. Your summaries often capture the essence of the original text."},
{"role": "user", "content": "I have a long transcript that I would like you to summarize for me. Please think carefully and do the best job you possibly can."},
{"role": "system", "content": "Absolutely, I will provide a concise and comprehensive summary of the transcript."},
{"role": "user", "content": "Excellent, here is the transcript: " + transcript_text}
],
max_tokens=3800 - token_len,
n=1,
stop=None,
temperature=0.5,
)
summary = response['choices'][0]['message']['content']
with open("transcript_summary.txt", "w") as file:
file.write(summary)
return summary.strip()```
While I've set up a system that tries to handle each stage sequentially using Streamlit's session_state, I'm encountering unexpected behavior.
- After transcribing, the summarization doesn't trigger. The app just refreshes and I get the message: "An error occurred: name 'summarized_text' is not defined" (this happens when uploading small mp3-files with less than 25MB)
- But when uploading large mp3 files (>25MB), I get the message: "An error occurred: name 'transcription' is not defined" right away.
I thing I messed up the states somehow and the chunking parts....
我会逐一回答：