Profanity Filtering in STT: Custom Post-Processing with pymorphy3

Imagine your platform processes audio chat for children. A user utters a profane word in the genitive case. The built-in Google STT filter misses it — no exact match. Result: complaints, bans, reputation damage. To avoid this, you need combined filtering: provider + morphological post-processing. We implement such a turnkey solution in 2–5 days. We have experience in 10+ commercial projects, processing up to 1000 hours of audio per day. Filtering accuracy is at least 95%. Meanwhile, the average budget savings on moderation is 60%.

Why Providers Can't Handle It Alone?

Built-in filters of Google, AWS, and Azure are simple but have limitations. Let's compare them:

The table shows that none account for morphology. For Russian this is critical: a word can be in any grammatical form. For example, a profane word in the genitive case will pass through the provider's filter if there is no exact match. Therefore, we add post-processing based on pymorphy3.

Comparison of Filtering Methods

Our approach is 3 times more accurate compared to direct substring search (verified on our benchmark). According to pymorphy3 documentation, lemmatization ensures accuracy over 95%.

How Filtering Solves Legal Requirements?

For platforms with child content or corporate systems, filtering is not only ethics but also law. GDPR and 152-FZ require protection of minors from harmful content. Automatic filtering replaces manual moderation, reducing costs by 60% and eliminating human error. We configure logging so that only trigger labels are stored — no audio or transcription is saved.

How Does Morphological Post-Processing Work?

We use Azure Speech Profanity filter as a base, and on top we apply our Python module. Example code:

import pymorphy3

morph = pymorphy3.MorphAnalyzer()
PROFANITY_SET = {"badword1", "badword2", "badword3"}  # normal forms

def filter_text(text: str, replacement: str = "***") -> str:
    result = []
    for token in text.split():
        norm = morph.parse(token)[0].normal_form
        if norm in PROFANITY_SET:
            result.append(replacement)
        else:
            result.append(token)
    return " ".join(result)

What's Included in the Work?

• Audit of the current STT system and filtering requirements.
• Configuration of the provider (Google, AWS, Azure) with built-in filter enabled.
• Development and integration of the post-processing module in Python with pymorphy3.
• Expansion of the profanity dictionary for your content.
• Testing on a representative sample (minimum 1000 phrases).
• Documentation for setup and operation.
• Training of your team.
• Two-week support after implementation.

Step-by-Step Implementation Process

1. Analysis of current stack and filtering requirements (languages, audio volume, needed accuracy).
2. Configuration of the STT provider with built-in filter enabled.
3. Development and integration of the post-processing module with pymorphy3.
4. Expansion of the profanity dictionary based on your data.
5. Testing on 10+ audio files with different grammatical forms.
6. Documentation and two-week support.

Timeline: 2 to 5 business days. Cost is calculated individually — a typical project pays for itself in 2 months through reduced manual moderation.

Common Mistakes and How to Avoid Them

• Using only regex — misses modifications (emojis, letter substitutions). Accuracy drops to 60%.
• Relying solely on the provider — does not cover rare profanities. Example: a word in the instrumental case is missed.
• Not updating the dictionary — new words appear every 3–6 months. Need automated monitoring.
• Logging content — violates law: store only the fact of a trigger and a timestamp.

How Is the Filter Tested?

We run 1000 audio files with known annotations. We measure Precision and Recall at the token level. Target metrics: Precision > 98%, Recall > 95%. If not achieved, we refine the dictionary or replacement rules. Result: p99 latency < 200 ms.

Contact us for an audit of your current system — we will offer the optimal solution. Get a consultation on implementing filtering today! Order the implementation of profanity filtering in STT.

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.