Development of an AI Transcription System with API
You have a call analytics startup, and you integrated a ready-made transcription API — within a month the cost per minute of audio tripled, and the p99 latency exceeded 12 seconds. A typical situation: third-party services don't scale to your volume, and their pricing is uncontrollable. We build private transcription APIs — with full control over the model, infrastructure, and cost. With experience in ML and 40+ deployed systems, we guarantee stable performance under any load.
Why Deploy Your Own Transcription API?
Off-the-shelf solutions fall short when you need to process hundreds of hours of audio per day, achieve sub-5-second latency, or customize the model for your domain. A custom API built on Whisper with diarization and batching delivers p99 < 4 seconds even on an A10G. The operational cost per minute is 5–10 times lower than cloud providers. Our API is 4 times faster than cloud counterparts in p99 for transcribing one minute of audio.
What Problems We Solve
Latency. During audio calls, the client waits for the transcript to search for keywords. Our API streams results via WebSocket with < 3 seconds delay on a 30-second chunk. For comparison, cloud APIs typically have p99 of 8–15 seconds — at least 4 times longer.
Diarization quality. Standard APIs confuse speakers if voice timbres are similar. We use pyannote-audio with a pre-trained embedder — diarization accuracy of 92% on the CorpSet.
Billing integration. You need to charge for each minute of transcription. The API includes a built-in consumption counter with automatic limits and notifications.
How We Do It
Stack and Architecture
- Model:
Whisper Large V3 (INT8 quantized) inferred via TensorRT — up to 2x speedup on the same GPU.
- Service:
FastAPI asynchronously (Uvicorn + Gunicorn). Task queue on Celery with Redis. For streaming — WebSocket via WebSockets.
- Diarization:
pyannote-audio in a separate container, results merged by timestamps.
- Batching:
vLLM for Whisper — groups up to 32 audio chunks into one batch, smoothing latency.
Real-World Case
A call center analytics platform was processing 4000 hours of audio per month via a cloud API — after deployment, the budget was cut 5 times. We deployed a private API on 2× L40S: p99 dropped from 15 s to 3.2 s. Stack: Whisper + pyannote + TensorRT + Celery. We implemented batching and result caching for repeated requests.
How We Ensure High Availability
The architecture is built on Kubernetes with automatic horizontal scaling: as load increases, inference pods are added. Monitoring via Prometheus + Grafana, alerts on p99 latency and GPU utilization. 99.9% uptime is contractually guaranteed.
Process of Work
-
Analysis: We break down your requirements — volume, latency, model, CRM integration.
-
Design: API specification (REST + WebSocket), inference server choice (vLLM / TGI / Triton), billing scheme.
-
Development: Implementation of endpoints, diarization, batching, SDK (Python/JS).
-
Testing: Load testing with your data — measuring p99, FLOPS, GPU utilization.
-
Deployment: Containerization (Docker + Kubernetes), monitoring (Prometheus + Grafana), documentation (OpenAPI + Postman).
What's Included
- REST API + WebSocket endpoints (as per specification above)
- Webhook notifications for task status
- SDK for Python and JavaScript
- Per-minute billing with limits
- OpenAPI documentation + Swagger UI
- 1 month of support after deployment
Comparison of Approaches
| Parameter |
Ready-Made Cloud API |
Custom API (Our Development) |
| latency p99 (1 min audio) |
8–15 s |
2–4 s |
| Cost / minute |
high (depends on provider) |
low (depends on load) |
| Model customization |
no |
full (fine-tuning, LoRA) |
| Data control |
under NDA |
your infrastructure |
| Scaling |
quotas |
automatic horizontal scaling |
Comparison of Inference Models
| Model |
Speed (latency p99) |
Quality (WER) |
GPU Memory |
| Whisper Large V3 (FP16) |
6.1 s |
8.2% |
10 GB VRAM |
| Whisper Large V3 (INT8) |
3.4 s |
8.5% |
5 GB VRAM |
| Distil-Whisper (FP16) |
1.8 s |
10.1% |
4 GB VRAM |
How We Optimize the Model for Your Data
If your corpus contains specialized vocabulary (medical, legal), we can fine-tune Whisper using LoRA. This reduces WER by an additional 5–15% without increasing latency. A labeled dataset of at least 10 hours is required.
Estimated Timeline
- Baseline API (REST + WebSocket, single model, diarization): 2 to 4 weeks.
- With billing, SDK, load testing: 1 to 2 months.
- From specification approval, depending on customization complexity.
We'll provide an accurate estimate after analyzing your requirements — contact us to discuss details within 1 day. Over 40 companies have entrusted us with their transcription systems. Request a free consultation with no obligation.
Original diarization research: pyannote-audio
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.
