Automatic Court Proceeding Transcription Implementation

Implementation of automatic transcription of court hearings Transcription of court hearings is a highly demanding task with accuracy requirements of >98% (WER <2%). Specific features: multiple speakers, interruptions, legal vocabulary, procedural formulas, proper names. The transcript is a procedural document. ### System requirements - WER <5% on legal vocabulary (after post-processing) - Accurate attribution of remarks (chairman, prosecutor, lawyer, witness, defendant) - Timestamps for each remark - Automatic normalization of numerals, dates, articles of law - Secure storage (data does not leave the circuit) ### On-premise architecture

# Полностью локальное развёртывание без облака
class CourtTranscriptionSystem:
    def __init__(self):
        # Whisper large-v3 дообученный на юридических данных
        self.stt = WhisperModel(
            "/models/whisper-legal-ru",
            device="cuda",
            compute_type="int8_float16"
        )
        # Диаризация — обязательно
        self.diarizer = Pipeline.from_pretrained(
            "pyannote/speaker-diarization-3.1"
        )
        # Нормализатор юридических текстов
        self.normalizer = LegalTextNormalizer()

    async def transcribe_session(self, audio_path: str, participants: dict) -> dict:
        """
        participants: {"SPEAKER_00": "Председатель Иванова И.И.", ...}
        """
        # 1. Транскрибируем с word timestamps
        segments, _ = self.stt.transcribe(
            audio_path,
            word_timestamps=True,
            language="ru",
            vad_filter=True,
            initial_prompt="Судебное заседание. Председатель суда, прокурор, адвокат, подсудимый."
        )

        # 2. Диаризация
        diarization = self.diarizer(audio_path)

        # 3. Сопоставляем с участниками
        labeled_transcript = self._label_speakers(
            list(segments), diarization, participants
        )

        # 4. Нормализация: "сто пятьдесят вторая статья" → "ст. 152"
        for segment in labeled_transcript:
            segment["text"] = self.normalizer.normalize(segment["text"])

        return {
            "session_date": datetime.now().isoformat(),
            "transcript": labeled_transcript,
            "metadata": {
                "audio_duration": self._get_duration(audio_path),
                "speaker_count": len(participants)
            }
        }
```### Specialized dictionary for the court```python
LEGAL_VOCABULARY = [
    "апелляционное определение",
    "постановление о прекращении дела",
    "кассационная жалоба",
    "статья двести шестьдесят четвёртая",
    "часть первая статьи",
    "Уголовно-процессуальный кодекс",
    "Гражданский процессуальный кодекс",
    # ... несколько тысяч терминов
]
```### Export to formats - **DOCX** with table: time | speaking | text - **XML** for court electronic document management systems - **PDF** with signature and hash for integrity ### Post-editing The system is designed to **support the secretary**, not replace: it produces a draft with >90% accuracy, the secretary corrects and certifies. ### Implementation time - Basic system: 4–6 weeks
- With additional training on legal data: +4–6 weeks - Certification and integration into the State Automated System “Justice”: a separate project