Deepgram integration for low-latency streaming STT

We integrate Deepgram for streaming speech recognition with latency under 200 ms. When your product needs real-time STT—live subtitles, voice assistants, call analytics—standard solutions like Google Speech-to-Text deliver 500-800 ms latency and require post-processing. Deepgram Nova-2 outputs results 60% faster with comparable WER of 5-8% on English and 12-18% on Russian (beta).

What problems does Deepgram solve?

High latency. Many cloud STT providers buffer audio and send it in packets, adding delay. Deepgram offers WebSocket streaming with immediate intermediate and final results. We design the architecture so p99 latency stays under 250 ms.

Russian language quality. For Russian, Deepgram is still beta with WER 12-18%, higher than English. To reduce error rates, we calibrate the language model (domain-specific tuning), add custom keywords, and apply rule-based post-processing.

Diarization and analytics. Determining "who spoke when" in multi-channel calls is nontrivial. Deepgram supports channel and speaker diarization but requires voice profile tuning. We integrate with call metadata.

Why Nova-2 outperforms base models

Nova-2 processes audio 30 times faster than real time (30x RT)—3-5x faster than the Base model with the same WER. This is achieved through an end-to-end attention architecture that eliminates separate decoding stages. For comparison: Google Chirp delivers 10x RT, Whisper large 3x RT. Deepgram wins on speed, sacrificing accuracy on rare languages.

How to reduce streaming latency

Key parameters:

• Use WebSocket instead of REST (REST adds up to 500 ms round-trip)
• Disable buffering options (e.g., punctuate=false, interim_results=true)
• Reduce chunk size to 10-20 ms (4096 bytes for 16 kHz audio)
• Choose the Base model if accuracy is not critical

In a live webinar transcoding project, we reduced latency from 600 ms to 180 ms using these optimizations.

How we integrate Deepgram: stack and approach

The basic scenario is WebSocket integration with a persistent connection. We use Python asyncio with the websockets library and the official Deepgram SDK for authentication. Example streaming code (Nova-2, Russian, diarization):

import asyncio
import websockets
import json

async def transcribe_stream():
    url = "wss://api.deepgram.com/v1/listen"
    headers = {"Authorization": f"Token {DEEPGRAM_API_KEY}"}
    params = "?model=nova-2&language=ru&punctuate=true&diarize=true"

    async with websockets.connect(url + params, extra_headers=headers) as ws:
        async def send_audio():
            with open("audio.wav", "rb") as f:
                while chunk := f.read(4096):
                    await ws.send(chunk)
            await ws.send(json.dumps({"type": "CloseStream"}))

        async def receive_results():
            async for message in ws:
                result = json.loads(message)
                if result.get("is_final"):
                    transcript = result["channel"]["alternatives"][0]["transcript"]
                    print(transcript)

        await asyncio.gather(send_audio(), receive_results())

We also configure parameters: utterances=true for phrase splitting, numerals=true for digits, smart_format=true for punctuation and symbols.

What the work includes

• Audit of current architecture—evaluate latency, language, and volume requirements.
• Integration design—select model, protocol (REST/WebSocket), authentication scheme.
• Implementation—write the module for your backend (Python, Node.js, Go, Java).
• Testing—A/B testing on a test dataset, measuring p50/p95/p99 latency.
• Documentation—API description, configs, deployment guide.
• Team training—workshop on operation and monitoring.

Process

1. Analytics (2-3 days)—gather requirements, select Deepgram model, evaluate quality on your audio.
2. Design (2-4 days)—develop architecture, agree on protocols and error handling (reconnect, backpressure).
3. Implementation (5-10 days)—WebSocket integration, diarization setup, post-processing.
4. Testing (3-5 days)—run on test data, optimize parameters, load test.
5. Deployment and handover (2-3 days)—deploy in your environment (AWS/GCP/on-prem), hand over documentation.

Timeline and cost

Timeline: 2 to 4 weeks, depending on complexity (streaming vs batch, custom model needs, diarization). Cost is calculated individually based on work volume and selected stack. We start with a free technical audit—evaluate your current architecture and provide a preliminary plan. Contact us to discuss your project. Get a consultation—we'll send details on timeline and cost.

Deepgram vs alternatives

Based on real-time speech experience, we guarantee integration with p99 latency under 300 ms and accuracy comparable to reference models. Our engineers hold Deepgram and AWS certifications, ensuring reliable solutions.

According to Deepgram, the Nova-2 model saves up to 40% in costs compared to Google STT under high loads. Contact us for a free audit of your project.

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.