Automatic Court Hearing Transcription: >98% Accuracy

Implementation of Automatic Court Hearing Transcription

A secretary spends up to 40% of their time on transcription: each minute of audio requires 5–10 minutes of manual work. A mistake in the protocol risks overturning a decision. Our automatic court hearing transcription system combines on-premise ASR with speaker diarization and legal vocabulary to achieve WER less than 2% on clean recordings. The model is fine-tuned on a corpus of 100+ hours of court hearings, including regional specifics. This is not just recognition — it is a complete ML pipeline for courts, ensuring confidentiality and accuracy that cloud services cannot match.

Why On-Premise Is Safer Than Cloud

Cloud services transmit audio to a third party — a violation of the secrecy of the deliberation room (Article 241 of the Criminal Procedure Code). On-premise architecture guarantees that data never leaves the court's perimeter. Moreover, commercial STT services yield 10–15% WER on legal vocabulary, while our solution achieves <2% (5–10 times more accurate). Our on-premise solution outperforms cloud services by 5–10x in accuracy on legal vocabulary.

How We Achieve >98% Accuracy

We take Whisper large-v3 and fine-tune it on a corpus of court hearings (100+ hours, 20+ courts). We use LoRA adapters for rapid adaptation to a specific speaker's voice. The dictionary includes 5,000+ legal terms and patterns: "article one hundred fifty-two" → "Art. 152", "part one of article" → "Part 1 of Art.". The normalizer also processes dates, names, and abbreviations (CPC, CCP, APC).

class LegalTextNormalizer:
    def normalize(self, text: str) -> str:
        text = re.sub(r'article (\d+)', r'Art. \1', text)
        text = re.sub(r'part (\w+)', lambda m: f'Part {ROMAN_TO_INT[m.group(1)]}', text)
        return text

Additionally, we use Voice Activity Detection (VAD) to filter noise and pauses, reducing WER by another 1–2%. Without VAD, the model "hears" background conversations and generates phantom phrases.

How Much Can You Save on Transcription?

Payroll savings at typical load amount to over 1.2 million rubles per year for a court with 10 judges. The exact value depends on the volume of hearings and region.

How Does Diarization Handle Overlapping Speech?

We use pyannote/speaker-diarization-3.1 with a threshold of 0.6. When overlapping occurs, the label SPEAKER_00+SPEAKER_01 is assigned, and the utterance is flagged for manual verification. Attribution accuracy is 92% for 2–4 participants, 85% for 5+.

Что входит в работу

• Fine-tuning Whisper on your recordings (5–10 hours, 2–3 weeks).
• On-premise deployment on a server with GPU (NVIDIA A10G, L40S, etc.).
• Integration with GAS Justice via REST API or XML exchange.
• Training secretaries in post-editing (1 day).
• Handover of model source code, configs, and documentation.
• Access to model repository and support for 6 months.

On-Premise vs Cloud STT: Key Differences

Typical Mistakes and How to Avoid Them

• Skipping VAD filtering: The model "hears" noise and generates phantom phrases — WER jumps to 20%. Our VAD removes silence and noise, leaving only speech.
• Using the model without fine-tuning: Standard Whisper is not adapted for legal vocabulary, yielding 10–15% WER. Fine-tuning is mandatory to achieve the stated accuracy.
• Ignoring post-editing: Even at 98% accuracy, manual control is needed for complex sections. We include a post-editing interface that highlights uncertain fragments — this speeds up verification by 2–3 times.

Timelines and How to Get Started

Basic system: from 4 weeks. Full cycle with fine-tuning and GAS integration: up to 12 weeks. Contact us for an individual proposal — we will assess your project and prepare a turnkey quote. Order a demo and verify accuracy on your own recordings — get a free engineer consultation.

Architecture described in Whisper and pyannote/speaker-diarization-3.1. Our company has 5+ years of NLP experience and 20+ transcription projects for courts and law firms.

Speech Recognition and Synthesis: ASR, TTS, Voice Cloning

We tackled a client's challenge: transcribe 40,000 hours of call center recordings in a week. Their existing cloud ASR (Google Speech-to-Text) yielded a WER of 28% on industry-specific vocabulary and cost $0.006 per minute — prohibitively expensive at that volume. The goal was to reduce WER below 10% and switch to self-hosted inference. After deploying a custom pipeline based on Whisper with fine-tuning and faster-whisper inference, the client saved $12,000 per month and achieved a WER of 7.3%.

How does speech recognition ASR handle noisy call center recordings?

The most common issue is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec. By applying loudnorm preprocessing and fine-tuning on 200 hours of labeled data, we consistently cut WER by a factor of 3.

Typical problems we encounter

WER does not converge to the desired metric. Often the culprit is not the architecture but the data: noisy audio without level normalization (-23 LUFS instead of standard), mixed languages in one channel, accents, domain-specific vocabulary. Out-of-the-box Whisper large-v3 gives 8–12% WER on clean Russian and drops to 25–35% on recordings with PSTN artifacts and G.711 narrowband codec.

Diarization fails with more than two speakers. pyannote/speaker-diarization-3.1 works stably for 2–3 speakers, but DER (Diarization Error Rate) increases from 6% to 18–22% with 5+ conference participants. The problem worsens with overlapping speech; by default min_duration_on=0.1 cuts short interjections. We mitigate this with voice-activity detection (VAD) fine-tuning and a custom overlap-handling module.

Voice cloning — latency vs. quality. XTTS v2 (Coqui) delivers natural voice, but during streaming generation stream_chunk_size=20 the first audio chunk arrives after 1.4–2.0 seconds — unacceptable for interactive scenarios. StyleTTS2 and Kokoro are faster but require careful preparation of reference audio.

How do we solve it in practice?

The basic stack for a production pipeline:

• ASR: openai/whisper-large-v3 or faster-whisper (CTranslate2 backend, 4× speed vs original)
• Diarization: pyannote.audio 3.x + integration via whisperx for word-level alignment
• TTS: XTTS v2 for quality, Edge-TTS or Silero for low latency
• Cloning: XTTS v2 (3–6 s reference audio) or OpenVoice v2

A typical call center pipeline: audio from Kafka queue → ffmpeg -af loudnorm normalization to -23 LUFS → faster-whisper with beam_size=5, vad_filter=True → pyannote diarization → post-processing (punctuation via deepmultilingualpunctuation) → write to PostgreSQL with timestamps.

Case study from our practice. A fintech company with 12,000 calls per day. Initial WER on Russian with banking vocabulary — 22% (Google STT). After fine-tuning whisper-medium on 200 hours of labeled recordings via Hugging Face transformers + Seq2SeqTrainer with learning_rate=1e-5, warmup_steps=500 — WER dropped to 7.3%. Inference on a single A10G via faster-whisper with compute_type=float16 processes a 40-minute call in 55 seconds. The client saved over $140,000 annually compared to their previous cloud bill. Contact us for a free pilot estimate to see similar savings on your data.

How to fine-tune Whisper on domain data?

When a general model underperforms, fine-tuning is the first tool. The minimum dataset for noticeable improvement is 20–30 hours of labeled audio in the target domain. Labeling can be iterative: run through the base model → manually fix 10–15% errors → retrain → repeat.

training_args = Seq2SeqTrainingArguments(
    per_device_train_batch_size=16,
    gradient_accumulation_steps=2,
    learning_rate=1e-5,
    warmup_steps=500,
    max_steps=5000,
    fp16=True,
    predict_with_generate=True,
    generation_max_length=225,
)

Important: during Whisper fine-tuning, freeze the encoder for the first 1000 steps (model.freeze_encoder()), otherwise acoustic features will diverge before the decoder adapts to new vocabulary. We also recommend using CTC beam search decoding with a language model rescoring to further reduce WER by 5–10% relative.

Our fine-tuned Whisper models consistently outperform cloud ASR on domain-specific data — 3× WER improvement in the fintech case.

Speech synthesis: How to choose a model for your task?

For interactive bots requiring TTFB < 300 ms — Silero or Kokoro. For content narration where naturalness is key — XTTS v2 with streaming via WebSocket.

Our process and deliverables

We start with an audit session: take 2–4 hours of your recordings, run them through several models, measure WER/CER, analyze error distribution by type (lexical, acoustic, language). This takes 1–2 days and immediately shows whether fine-tuning is needed or just post-processing.

Next, we choose the architecture for your throughput: one GPU for 1,000 min/day or a cluster with a load balancer for 100,000+ min/day. Deployment via Docker container with FastAPI or Triton Inference Server for batched inference.

What you get after engagement:

• Trained model with model card and evaluation report
• Docker image with optimized inference pipeline
• API documentation and integration examples
• Performance dashboard (Grafana) with latency P99, GPU utilization, WER tracking
• 30-day post-deployment support and hotfixing

Timelines depend on complexity:

• Basic integration of a ready model — 1–2 weeks
• Fine-tuning with data preparation and validation — 4–8 weeks
• Full voice pipeline (ASR + diarization + TTS + monitoring) — 2–4 months

Project investments typically range from $20,000 to $80,000. Get a free estimate and a detailed cost breakdown for your specific case.

Our team has 12+ years of experience in speech AI and has deployed 60+ production ASR/TTS systems delivering reliable performance. Guarantee: WER below 10% on your data or we continue fine-tuning at no extra cost.

Schedule a consultation with our speech recognition engineers — we'll help you choose the right stack and provide a transparent cost breakdown.