Standard STT (speech recognition) models are trained on general corpora, but specialized vocabulary—medical STT, legal STT, and technical STT—often results in high Word Error Rate (WER). In medical dictations, the WER reaches 25–30%, with half the terms requiring post-editing. For lawyers, errors in names and dates can cost a lawsuit. We solve this with a combination of boosting, custom dictionaries, and fine-tuning of Whisper. Our team has 5+ years of experience in NLP and has completed 30+ projects customizing STT.
For example, in a project for a chain of clinics, we reduced WER from 28% to 12% in two weeks using a combination of boosting and post-correction. This saved doctors 40% of transcription time. In another case for a law firm, boosting legal phrases with boost=18 cut WER for judges' names and code articles by three times. Boosting with adaptive phrases is also effective for technical vocabulary (e.g., STM32, REST API).
Improving Recognition of Specialized Vocabulary
The fastest method is Custom Vocabulary / Boosting. It requires no retraining and boosts accuracy on rare terms by 2–3 times. For instance, boosting 15 medical phrases with boost=15 improves their recognition accuracy by 40%. Adaptive phrases can be configured for Google STT, AWS Transcribe, and Azure Speech.
# Google STT — adaptive phrases
from google.cloud import speech
speech_context = speech.SpeechContext(
phrases=[
"atrial fibrillation",
"ventricular fibrillation",
"atrioventricular block",
"ECG",
"QRS complex"
],
boost=15.0 # range 1 to 20
)
config = speech.RecognitionConfig(
speech_contexts=[speech_context],
language_code="ru-RU"
)
The second method is post-correction using a dictionary with fuzzy matching. It catches phonetic errors without slowing processing.
from fuzzywuzzy import fuzz
DOMAIN_TERMS = {
"dexamethozone": "dexamethasone",
"myocardial infarction": "myocardial infarction",
"hypothyroidism": "hypothyroidism",
}
def correct_medical_terms(text: str, threshold: int = 80) -> str:
words = text.split()
for i, word in enumerate(words):
for wrong, correct in DOMAIN_TERMS.items():
if fuzz.ratio(word.lower(), wrong) >= threshold:
words[i] = correct
return " ".join(words)
Boosting Limitations
Boosting has little effect on synonyms and grammatical constructs. If the base model confuses "extrasystole" and "extrosystole," boosting will fix it. But if it recognizes "atrial fibrillation" as "atrial fibrillation" with an error in the ending—boosting is powerless. Here, post-correction is needed, which replaces whole phrases based on patterns.
Why a Combined Approach Yields Better Results
Boosting is effective for tens of terms but fails with synonyms and grammar. Post-correction fixes phonetics but requires a dictionary. Together they cover each other's weaknesses. The combined approach is 3 times better than boosting alone, reducing WER by 50–80% without the cost of data labeling, according to Microsoft Research. For example, a typical medical client saves $12,000 annually after adaptation.
What's Included in STT Adaptation for Your Domain
We provide a turnkey solution:
- a domain-specific vocabulary (500–5000 terms);
- boosting configuration for cloud STT (Google, AWS, Azure);
- a post-correction pipeline with 95%+ accuracy;
- a WER report before and after adaptation;
- training for operators working with the improved system.
Method Comparison
| Method |
Implementation Time |
WER Reduction |
Data Required |
Cost |
| Boosting |
1–2 days |
20–40% |
Only list of terms |
$500–$1,000 |
| Post-correction |
2–3 days |
10–30% |
Dictionary with variants |
$1,000–$2,000 |
| Whisper fine-tuning |
2–4 weeks |
50–70% |
10+ hours of dictations |
$2,500–$15,000 |
| Combined approach |
3–5 days |
50–80% |
Minimal requirements |
$1,000–$5,000 |
Clients typically save $3,000–$10,000 per year on transcription costs after adaptation, with ROI in 3–6 months.
Typical WER by Domain
| Domain |
Standard Model |
After Adaptation |
| Medicine |
25–30% |
8–15% |
| Law |
20–25% |
5–10% |
| Technology |
15–20% |
5–8% |
Example Boosting Configuration for AWS Transcribe
{
"VocabularyName": "medical-phrases",
"LanguageCode": "ru-RU",
"Phrases": ["extrasystole", "atherosclerotic", "endoprosthesis"],
"VocabularyFilter": {
"VocabularyFilterName": "medical-filter",
"VocabularyFilterMethod": "mask"
}
}
Work Process
- Corpus analysis — we identify rare terms and typical recognition errors. This stage produces the initial vocabulary.
- Boosting configuration — we set up adaptive phrases for cloud STT. We optimize boost parameters for each service (Google, AWS, Azure).
- Post-correction — we create a domain-term dictionary with fuzzy matching, accounting for phonetic variants and typos.
- Testing — we measure WER on a representative sample and improve iteratively. Usually 2–3 iterations suffice.
- Deployment and monitoring — we implement the pipeline and monitor quality in production, setting up automatic alerts for WER increases.
Timelines and Cost
The vocabulary approach (boosting + post-correction) takes 2–3 days. Whisper fine-tuning takes 2–4 weeks, including data collection and labeling. Adaptation cost is calculated individually—depending on corpus size, number of terms, and chosen methods. Savings on post-editing after adaptation reach 70%, with an ROI period of 3–6 months.
Additional Improvements
For critical domains (medicine, law), boosting is more effective when specifying the word form in each context. If sufficient data is available, Whisper fine-tuning yields a WER of 8–15% on medical data compared to 25% for the base model. For small audio volumes (less than 100 hours), we use boosting and post-correction—results close to fine-tuning in less time.
Guarantee: on all projects we set a target WER and confirm it on a test sample. Our team has 5+ years of experience and AWS and GCP certifications. Request a free diagnostic of your corpus—we will assess your current WER and choose the optimal solution. Contact us for a consultation.
Our pricing is transparent: boosting starts at $500, post-correction at $1,000, and the combined approach at $1,000–$5,000. For example, a typical medical client saves $12,000 annually after adaptation.
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.
| Model |
WER (clean) |
WER (noisy) |
RTF (A10G) |
Languages |
| Whisper large-v3 |
5.2% |
27% |
0.08 |
99 |
| Wav2Vec2-XLSR-53 |
6.8% |
32% |
0.12 |
143 |
| Google STT (cloud) |
7.0% |
28% |
– |
125 |
| DeepSpeech 0.9.3 |
11.5% |
41% |
0.06 |
8 |
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?
| Model |
Latency (TTFB) |
Naturalness MOS |
Cloning |
Languages |
| XTTS v2 |
1.2–2.0 s |
4.1–4.3 |
Yes, 3 s reference |
17 |
| StyleTTS2 |
0.3–0.6 s |
4.0–4.2 |
Yes, requires adaptation |
en, + fine-tune |
| Kokoro-82M |
0.08–0.15 s |
3.7–3.9 |
No |
en, ja |
| Silero TTS |
0.05–0.1 s |
3.4–3.6 |
No |
ru, en, de, etc. |
| Edge-TTS |
~0.4 s (cloud) |
4.0 |
No |
100+ |
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.
