vikarti.anatra/AudioSumma
1
0
Fork 0
mirror of https://github.com/patw/AudioSumma.git synced 2025-09-04 19:50:29 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Used mistral large to re-write the summarize process.

Browse source
This commit is contained in:
Pat Wendorf 2024-08-19 07:41:32 -04:00
parent 81e4e165bc
commit 3c05cd78dd
1 changed files with 38 additions and 72 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

110
summarize.py
View file

@ -21,23 +21,18 @@ CHUNK_SIZE = int(os.getenv("CHUNK_SIZE"))
TEMPERATURE = float(os.getenv("TEMPERATURE")) TEMPERATURE = float(os.getenv("TEMPERATURE"))
def whisper_api(file): def whisper_api(file):
# Whisper supports multiple files, but we're sending one """Transcribe audio file using Whisper API."""
files = {"file": file} files = {"file": file}
# Required API call data
api_data = { api_data = {
"temperature": "0.0", "temperature": "0.0",
"response_format": "json" "response_format": "json"
} }
# Call API and return text
response = requests.post(WHISPERCPP_URL, data=api_data, files=files) response = requests.post(WHISPERCPP_URL, data=api_data, files=files)
return response.json()["text"] return response.json()["text"]
def llama_api(prompt): def llama_api(prompt):
# Format prompt before sending """Generate response using llama.cpp server API."""
formatted_prompt = PROMPT_FORMAT.format(system=SYSTEM_MESSAGE, prompt=prompt) formatted_prompt = PROMPT_FORMAT.format(system=SYSTEM_MESSAGE, prompt=prompt)
api_data = { api_data = {
"prompt": formatted_prompt, "prompt": formatted_prompt,
"n_predict": -1, "n_predict": -1,
@ -45,19 +40,15 @@ def llama_api(prompt):
"stop": [STOP_TOKEN], "stop": [STOP_TOKEN],
"tokens_cached": 0 "tokens_cached": 0
} }
response = requests.post(LLAMACPP_URL, headers={"Content-Type": "application/json"}, json=api_data) response = requests.post(LLAMACPP_URL, headers={"Content-Type": "application/json"}, json=api_data)
json_output = response.json() json_output = response.json()
return json_output['content'] return json_output['content']
# Use ffmpeg to trim silence in wav files, to prevent issues with
# whisper.cpp stopping the transcode if it detects a large amount of silence
def trim_silence(filename): def trim_silence(filename):
# Create a temporary file for the output """Trim silence from audio file using FFmpeg."""
with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_file: with tempfile.NamedTemporaryFile(suffix=".wav", delete=False) as temp_file:
temp_filename = temp_file.name temp_filename = temp_file.name
# Construct the FFmpeg command
ffmpeg_command = [ ffmpeg_command = [
"ffmpeg", "ffmpeg",
"-i", filename, "-i", filename,
@ -66,84 +57,59 @@ def trim_silence(filename):
temp_filename temp_filename
] ]
# Run the FFmpeg command
result = subprocess.run(ffmpeg_command, capture_output=True, text=True, check=True) result = subprocess.run(ffmpeg_command, capture_output=True, text=True, check=True)
# If FFmpeg command was successful, replace the original file
os.replace(temp_filename, filename) os.replace(temp_filename, filename)
# Iterate over each WAV file and transcode with whisper API def process_wav_files():
wav_files = [f for f in os.listdir(".") if f.endswith(".wav")] """Process WAV files: trim silence and transcribe."""
for wav_file in wav_files: wav_files = [f for f in os.listdir(".") if f.endswith(".wav")]
for wav_file in wav_files:
print("Trimming silence: " + wav_file)
trim_silence(wav_file)
# Trim silence on the wav file first with open(wav_file, "rb") as file:
print("Trimming silence: " + wav_file) print("Transcribing: " + wav_file)
trim_silence(wav_file) output_text = whisper_api(file)
output_file = os.path.splitext(wav_file)[0] + ".tns"
# Open the WAV file for sending to whisper REST API with open(output_file, "w") as output:
with open(wav_file, "rb") as file: output.write(output_text)
print("Transcribing: " + wav_file)
# Call whisper API to transcode file
output_text = whisper_api(file)
# Generate the output file name by replacing the extension with .tns
output_file = os.path.splitext(wav_file)[0] + ".tns"
# Write the output text to the file
with open(output_file, "w") as output:
output.write(output_text)
# Chunk the full transcript into multiple parts to fit in the context window
# and allow for better reasoning capability
def chunk_transcript(string, chunk_size): def chunk_transcript(string, chunk_size):
"""Chunk the transcript to fit in the context window."""
chunks = [] chunks = []
lines = string.split("\n") # Split the string on newline characters lines = string.split("\n")
current_chunk = "" current_chunk = ""
for line in lines: for line in lines:
current_chunk += line # Build up the string until the chunk size is reached current_chunk += line
if len(current_chunk) >= chunk_size: if len(current_chunk) >= chunk_size:
chunks.append(current_chunk) chunks.append(current_chunk)
current_chunk = "" current_chunk = ""
if current_chunk: # Add the last chunk if it's not empty if current_chunk:
chunks.append(current_chunk) chunks.append(current_chunk)
return chunks return chunks
# Get the current date in yyyymmdd format def summarize_transcripts():
today = datetime.datetime.now().strftime('%Y%m%d') """Summarize transcript files."""
today = datetime.datetime.now().strftime('%Y%m%d')
summary_filename = "summary-" + today + ".md"
transcript_files = [f for f in os.listdir(".") if f.endswith(".tns")]
# Modify the filename by appending the current date for transcript in transcript_files:
summary_filename = "summary-" + today + ".md" print("Summarizing: " + transcript)
with open(transcript, "r") as file:
transcript_data = file.read()
chunked_data = chunk_transcript(transcript_data, CHUNK_SIZE)
# Get the list of transcript files in the current directory
transcript_files = [f for f in os.listdir(".") if f.endswith(".tns")]
# Iterate over each WAV file
for transcript in transcript_files:
print("Summarizing: " + transcript)
# Open the WAV file
with open(transcript, "r") as file:
transcript_data = file.read()
# chunk the transcript so we don't blow out the context window
chunked_data = chunk_transcript(transcript_data, CHUNK_SIZE)
# Iterate through the chunks, and summarize them
for i, chunk in enumerate(chunked_data):
with open(summary_filename, "a") as md_file: with open(summary_filename, "a") as md_file:
# Generate call summary for i, chunk in enumerate(chunked_data):
summary_prompt = SUMMARY_PROMPT.format(chunk=chunk) summary = llama_api(SUMMARY_PROMPT.format(chunk=chunk))
summary = llama_api(summary_prompt) facts = llama_api(FACT_PROMPT.format(chunk=chunk))
sentiment = llama_api(SENTIMENT_PROMPT.format(chunk=chunk))
# Generate fact summary md_file.write(f"# Call Transcript - {transcript} - Part {i + 1}\n\nSummary: {summary}\n\nFacts:\n{facts}\n\nSentiment: {sentiment}\n\n---\n")
fact_prompt = FACT_PROMPT.format(chunk=chunk)
facts = llama_api(fact_prompt)
# Generate call sentiment print("Summarizing complete")
sentiment_prompt = SENTIMENT_PROMPT.format(chunk=chunk)
sentiment = llama_api(sentiment_prompt)
# Write the notes if __name__ == "__main__":
md_file.write(f"# Call Transcript - {transcript} - Part {i + 1}\n\nSummary: {summary}\n\nFacts:\n{facts}\n\nSentiment: {sentiment}\n\n---\n") process_wav_files()
summarize_transcripts()
print("Summarizing complete")