File size: 17,541 Bytes
5f0a430 6f46074 ced1fa1 6f46074 5f0a430 6f46074 5f0a430 6f46074 5f0a430 6f46074 5f0a430 6f46074 5f0a430 6f46074 5f0a430 6f46074 ced1fa1 6f46074 ced1fa1 6f46074 ced1fa1 6f46074 ced1fa1 6f46074 ced1fa1 6f46074 ced1fa1 6f46074 ced1fa1 6f46074 ced1fa1 6f46074 5f0a430 47b639c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 |
import logging
import os
import tempfile
from nltk.tokenize import sent_tokenize
import pickle
import re
from utils import clean_transcript, consolidate_similar_items, chunk_text, load_labels
import config
from models import load_diarization
import wave
import gc
import torch
import time
logger = logging.getLogger(__name__)
_diarize_model = None
def process_transcription(audio_content: bytes, whisper_model, enable_diarization=False):
start = time.time()
if not whisper_model:
raise ValueError("Whisper model not loaded.")
temp_file_path = None
try:
with tempfile.NamedTemporaryFile(delete=False, suffix=".wav") as temp_file:
temp_file_path = temp_file.name
temp_file.write(audio_content)
segments_gen, info = whisper_model.transcribe(temp_file_path, beam_size=5)
segments = list(segments_gen)
transcript = " ".join([seg.text.strip() for seg in segments])
global _diarize_model
if not enable_diarization:
return transcript, info, None
if _diarize_model is None:
_diarize_model = load_diarization(config)
if _diarize_model is None:
logger.warning("Diarization model not available, returning transcript without speakers")
return transcript, info, None
with wave.open(temp_file_path, 'rb') as wav:
frames = wav.getnframes()
rate = wav.getframerate()
#calcul audio duration
audio_duration = frames / float(rate)
if audio_duration < 3.0:
logger.info(f"Audio too short ({audio_duration:.2f}s), skipping diarization")
diarized_segments = [{"speaker": "SPEAKER_0", "text": transcript}]
diarized_transcript = f"[SPEAKER_0]: {transcript}"
return diarized_transcript, info, diarized_segments
logger.info("Running speaker diarization")
diarization = _diarize_model(temp_file_path)
# Extract diarization segments
diarize_segments = []
for turn, _, speaker in diarization.itertracks(yield_label=True):
diarize_segments.append({
"speaker": f"SPEAKER_{speaker.replace('SPEAKER_', '')}",
"start": turn.start,
"end": turn.end
})
diarized_segments = []
for segment in segments:
# Find best matching speaker based on time overlap
best_speaker = None
max_overlap = 0
seg_start = segment.start
seg_end = segment.end
for diar_seg in diarize_segments:
diar_start = diar_seg["start"]
diar_end = diar_seg["end"]
# Calculate overlap
overlap_start = max(seg_start, diar_start)
overlap_end = min(seg_end, diar_end)
if overlap_end > overlap_start:
overlap = overlap_end - overlap_start
if overlap > max_overlap:
max_overlap = overlap
best_speaker = diar_seg["speaker"]
# If no overlap found, assign to the closest speaker
if best_speaker is None:
min_distance = float('inf')
for diar_seg in diarize_segments:
# Distance to start of segment
dist_start = abs(seg_start - diar_seg["start"])
# Distance to end of segment
dist_end = abs(seg_end - diar_seg["end"])
# Take the minimum
dist = min(dist_start, dist_end)
if dist < min_distance:
min_distance = dist
best_speaker = diar_seg["speaker"]
diarized_segments.append({
"speaker": best_speaker or "SPEAKER_UNKNOWN",
"text": segment.text,
"start": segment.start,
"end": segment.end
})
# Format diarized transcript
diarized_transcript = ""
current_speaker = None
for segment in diarized_segments:
speaker = segment["speaker"]
text = segment["text"].strip()
if not text:
continue
if speaker != current_speaker:
diarized_transcript += f"\n[{speaker}]: {text}"
current_speaker = speaker
else:
diarized_transcript += f" {text}"
# Clean up memory
gc.collect()
if torch.cuda.is_available():
torch.cuda.empty_cache()
end = time.time()
logger.info("time : ", (end - start) * 10**3)
return diarized_transcript, info, diarized_segments
finally:
if temp_file_path and os.path.exists(temp_file_path):
os.remove(temp_file_path)
def process_summary(text: str, summarizer_pipeline, nlp_spacy, config):
if not summarizer_pipeline:
raise ValueError("Summarizer model not loaded.")
#clean transcript
cleaned_transcript = clean_transcript(text)
processed_text = cleaned_transcript
doc = None
if nlp_spacy:
try:
doc = nlp_spacy(processed_text)
sentences = [sent.text.strip() for sent in doc.sents]
processed_text = " ".join(sentences)
except Exception as e:
logger.error(f"SpaCy processing failed: {e}", exc_info=True)
categories = {
"meeting_title": [],
"intro": [],
"topics": [],
"decisions": [],
"action_items": [],
"questions": [],
"deadlines": [],
"participants": [],
"overall_summary": [],
"conclusion": []
}
# extraction meeting title
title_pattern = r'(meeting|call|session|discussion) (about|on|for|regarding) ([^.]+)'
title_matches = re.findall(title_pattern, processed_text, re.IGNORECASE)
if title_matches:
categories["meeting_title"].append(title_matches[0][2].strip())
if doc:
sentences = [sent.text.strip() for sent in doc.sents]
else:
try:
with open("/home/heymouad/nltk_data/tokenizers/punkt/english.pickle", "rb") as f:
tokenizer = pickle.load(f)
sentences = tokenizer.tokenize(processed_text)
except Exception as e:
logger.error(f"NLTK tokenization failed: {e}", exc_info=True)
sentences = sent_tokenize(processed_text)
# Find participants
people = set()
if doc:
for ent in doc.ents:
if ent.label_ == "PERSON":
person = ent.text.strip()
if len(person) > 2:
people.add(person)
if people:
categories["participants"] = list(people)
try:
# chunked the text because of limits of bart model
logger.info(processed_text[::100])
processed_text = chunk_text(processed_text)
parts_summaries = []
for chunk in processed_text:
result = summarizer_pipeline(chunk, max_length=150, min_length=30, do_sample=False)
if result and isinstance(result, list) and len(result) > 0:
part_summary = result[0].get('summary_text', '')
if part_summary:
parts_summaries.append(part_summary)
overall_summary = " ".join(parts_summaries)
overall_summary = summarizer_pipeline(overall_summary, max_length=150, min_length=30, do_sample=False)[0]['summary_text']
categories["overall_summary"] = [overall_summary]
except Exception as e:
logger.error(f"Summarization failed: {e}", exc_info=True)
categories["overall_summary"] = ["Failed to generate overall summary."]
# Process each sentence
for i, sentence in enumerate(sentences):
sentence = sentence.strip()
if not sentence:
continue
# Check for action items
if (re.search(r'(need to|will|shall|must|should|have to|assigned to|responsible for|task|action item|to-do|follow up|take care of)',
sentence, re.IGNORECASE) and
re.search(r'(we|you|I|they|he|she|team|group|department)', sentence, re.IGNORECASE)):
categories["action_items"].append(sentence)
continue
# Check for decisions
if re.search(r'(decided|agreed|conclusion|resolved|approved|rejected|consensus|finalized|confirmed|determined)',
sentence, re.IGNORECASE):
categories["decisions"].append(sentence)
continue
# Check for deadlines/timing with stronger patterns
if re.search(r'(by|due|deadline|schedule|date|tomorrow|next week|month|calendar|remind|upcoming|on|at|until)',
sentence, re.IGNORECASE) and re.search(r'(time|day|week|month|year|hour|minute)', sentence, re.IGNORECASE):
categories["deadlines"].append(sentence)
continue
# Check for questions/issues
if (re.search(r'(\?|issue|problem|concern|question|clarif|wonder|how|what|when|where|why|who)',
sentence, re.IGNORECASE) and
not re.search(r'(answer|answered|resolved|solved)', sentence, re.IGNORECASE)):
categories["questions"].append(sentence)
continue
# Check for intro statements
if i < len(sentences) // 10: # First 10% of sentences
if re.search(r'(welcome|begin|start|agenda|today|discuss|meeting|introduce|opening|good morning|hello|topic)',
sentence, re.IGNORECASE):
categories["intro"].append(sentence)
continue
# Check for conclusion statements
if i > len(sentences) * 9 // 10: # Last 10% of sentences
if re.search(r'(conclude|end|wrap|summary|thank|next meeting|follow up|adjourn|goodbye|bye|closing)',
sentence, re.IGNORECASE):
categories["conclusion"].append(sentence)
continue
# Everything else is considered a topic if it has substance
if len(sentence.split()) > 3: # Avoid very short sentences
categories["topics"].append(sentence)
# Process categories to avoid repetition and consolidate related points
for category in categories:
if category in ["topics", "action_items", "decisions", "questions", "deadlines"]:
categories[category] = consolidate_similar_items(categories[category])
# Limit the number of topics to avoid overwhelming
if len(categories["topics"]) > 10:
# If we have a summarizer, try to generate a summary of topics
try:
topics_text = " ".join(categories["topics"])
topics_summary = summarizer_pipeline(topics_text, max_length=200, min_length=50, do_sample=False)[0]['summary_text']
categories["topics"] = sent_tokenize(topics_summary)
except Exception as e:
logger.error(f"Topics summarization failed: {e}", exc_info=True)
# Otherwise just take the first few and last few topics
categories["topics"] = categories["topics"][:5] + categories["topics"][-5:]
# Add emojis to formatted output
formatted_summary = []
# Format meeting title if available
if categories.get("meeting_title"):
formatted_summary.append(f"π **Meeting Title:** {categories['meeting_title'][0]}")
formatted_summary.append("")
# Add overall summary
if categories.get("overall_summary"):
formatted_summary.append("π **Executive Summary:**")
formatted_summary.append(categories["overall_summary"][0])
formatted_summary.append("")
# Format participants
if categories["participants"]:
formatted_summary.append("π₯ **Participants:**")
formatted_summary.append(", ".join(categories["participants"]))
formatted_summary.append("")
# Format intro
if categories["intro"]:
formatted_summary.append("π― **Meeting Introduction:**")
formatted_summary.append(" ".join(categories["intro"]))
formatted_summary.append("")
# Format main topics
if categories["topics"]:
formatted_summary.append("π‘ **Key Topics:**")
for i, topic in enumerate(categories["topics"], 1):
formatted_summary.append(f"{i}. {topic}")
formatted_summary.append("")
# Format decisions
if categories["decisions"]:
formatted_summary.append("β
**Decisions Made:**")
for decision in categories["decisions"]:
formatted_summary.append(f"β’ {decision}")
formatted_summary.append("")
# Format action items
if categories["action_items"]:
formatted_summary.append("π **Action Items:**")
for item in categories["action_items"]:
formatted_summary.append(f"β’ {item}")
formatted_summary.append("")
# Format questions
if categories["questions"]:
formatted_summary.append("β **Questions & Concerns:**")
for question in categories["questions"]:
formatted_summary.append(f"β’ {question}")
formatted_summary.append("")
# Format deadlines
if categories["deadlines"]:
formatted_summary.append("β° **Deadlines & Timing:**")
for deadline in categories["deadlines"]:
formatted_summary.append(f"β’ {deadline}")
formatted_summary.append("")
# Format conclusion
if categories["conclusion"]:
formatted_summary.append("π **Conclusion:**")
formatted_summary.append(" ".join(categories["conclusion"]))
return "\n".join(formatted_summary)
def create_enhanced_summary_prompt(transcript: str, language_name: str) -> str:
"""
Creates a single, dynamic and insistent prompt that instructs the AI
to output its findings in the specified language.
"""
return f"""
You are an expert AI assistant. Your task is to analyze the following meeting transcript and extract key information into a structured JSON object.
**Primary Goal:** Analyze the provided transcript and generate a structured summary.
**CRITICAL LANGUAGE INSTRUCTION:** All text in your final JSON response must be written in the following language: **{language_name}**. There are no exceptions.
**ANALYSIS INSTRUCTIONS:**
1. Read the entire transcript to understand its context.
2. Identify a concise title for the meeting.
3. Identify all participants mentioned.
4. Write a brief paragraph summarizing the core themes and outcomes.
5. List all clear and agreed-upon decisions.
6. Extract all clear action items, identifying the task, who it was assigned to, the due date if mentioned, and the context.
**OUTPUT INSTRUCTIONS:**
- Respond ONLY with a valid JSON object.
- The JSON must use these exact keys: "meeting_title", "participants", "meeting_summary", "decisions_made", "action_items".
- **Language Check:** Before you finalize your response, verify that every single string value within the JSON is written in **{language_name}**.
**TRANSCRIPT TO ANALYZE:**
\"\"\"
{transcript}
\"\"\"
"""
def format_summary_to_markdown(summary_json: dict, language_code: str = "en") -> str:
"""Converts the structured JSON summary using labels from language files."""
# Load the correct set of labels
labels = load_labels(language_code)
summary_data = {k.lower().replace(" ", "_"): v for k, v in summary_json.items()}
# #### Use the loaded labels for all user facing text
if not summary_data.get("meeting_summary") and not summary_data.get("decisions_made") and not summary_data.get("action_items"):
return labels["no_content"]
markdown_parts = []
if title := summary_data.get("meeting_title"):
markdown_parts.append(f"### {title}\n")
if summary := summary_data.get("meeting_summary"):
markdown_parts.append(labels["summary_header"])
markdown_parts.append(summary)
markdown_parts.append("")
if decisions := summary_data.get("decisions_made"):
markdown_parts.append(labels["decisions_header"])
for decision in decisions:
markdown_parts.append(f"- {decision}")
markdown_parts.append("")
if action_items := summary_data.get("action_items"):
markdown_parts.append(labels["actions_header"])
for item in action_items:
task = item.get('task', 'N/A')
assigned_to = item.get('assigned_to', labels["not_specified"])
due_date = item.get('due_date', labels["not_specified"])
context = item.get('context', '')
markdown_parts.append(f"- **{labels['task_label']}**: {task}")
markdown_parts.append(f" - **{labels['assigned_to_label']}**: {assigned_to}")
markdown_parts.append(f" - **{labels['due_date_label']}**: {due_date}")
if context:
markdown_parts.append(f" - **{labels['context_label']}**: {context}")
markdown_parts.append("")
return "\n".join(markdown_parts) |