Real-Time Live Captions: Architecture, Latency, Integration

Real-Time Live Captions: Architecture, Latency, Integration

Real-time captions (live captions) are a technical rehabilitation tool per WCAG 2.1 (criterion 1.2.4) and the equivalent Russian standard GOST R 52872-2019. We build captioning systems that operate with less than 2 seconds of delay. This is critical for broadcasts, conferences, television, and educational platforms. Our team — 12 engineers with a combined experience of over 25 years in STT and NLP. We have implemented 10+ installations for events with audiences up to 5,000 people and provided information access for thousands of hearing-impaired users. Deploying AI captions can save up to 40% budget compared to manual captioning, with payback within 2–3 months for regular broadcasts. To assess your project, contact us — we will offer a turnkey solution.

Problems We Solve

Standard captions often lag 5–10 seconds behind speech — unacceptable for the hard of hearing. Typical challenges:

• Text-audio synchronization suffers when using batch audio processing.
• Cloud STT services don't always handle domain-specific vocabulary (medical, legal terminology).
• Integration with platforms like Zoom and Teams requires a separate bot and API setup.

We solve these by choosing streaming models, optimizing buffering, and customizing vocabulary. For instance, in one telemedicine conference project we fine-tuned Whisper on a medical term corpus — recognition accuracy rose from 82% to 95%.

How We Achieve Under 2 Seconds Latency

The key is choosing a streaming model and audio transmission architecture. Deepgram Nova-2 delivers partial results every 200ms, giving end-to-end latency around 1 second — 2–3 times faster than faster-whisper large-v3 in batch mode. For on-premise scenarios we use faster-whisper with a VAD filter and 3-second buffer, delivering 2.5–4 seconds. But if under 2 seconds is required — only cloud streaming.

Real-Time STT Stack

For captions with <2s latency from speech onset we use local faster-whisper or cloud Deepgram Nova-2. Core engine example in Python:

import asyncio
import websockets
from faster_whisper import WhisperModel
import numpy as np
import sounddevice as sd

class RealTimeCaptioner:
    def __init__(self):
        self.model = WhisperModel(
            "large-v3",
            device="cuda",
            compute_type="float16"
        )
        self.buffer = []
        self.chunk_duration = 3.0  # seconds of buffering
        self.sample_rate = 16000

    async def stream_captions(self, websocket, audio_queue: asyncio.Queue):
        """Stream captions via WebSocket"""
        while True:
            chunk = await audio_queue.get()
            self.buffer.append(chunk)

            buffer_duration = len(self.buffer) * len(chunk) / self.sample_rate

            if buffer_duration >= self.chunk_duration:
                audio_data = np.concatenate(self.buffer)
                self.buffer = []

                segments, _ = self.model.transcribe(
                    audio_data,
                    language="ru",
                    vad_filter=True,
                    vad_parameters={"min_silence_duration_ms": 500}
                )

                for segment in segments:
                    caption = {
                        "text": segment.text.strip(),
                        "start": segment.start,
                        "end": segment.end,
                        "confidence": segment.avg_logprob
                    }
                    await websocket.send(json.dumps(caption, ensure_ascii=False))

WebRTC Integration for Browser

The client side in JavaScript captures audio from the microphone and sends it to the server via WebSocket. The server returns captions, displayed with a rolling window.

// Client side: audio capture and streaming to server
class LiveCaptionClient {
    constructor(wsUrl) {
        this.ws = new WebSocket(wsUrl);
        this.captionDiv = document.getElementById('captions');
    }

    async startCapturing() {
        const stream = await navigator.mediaDevices.getUserMedia({
            audio: { sampleRate: 16000, channelCount: 1, echoCancellation: true }
        });

        const audioContext = new AudioContext({ sampleRate: 16000 });
        const processor = audioContext.createScriptProcessor(4096, 1, 1);

        processor.onaudioprocess = (event) => {
            const pcmData = event.inputBuffer.getChannelData(0);
            const int16Array = new Int16Array(pcmData.length);
            for (let i = 0; i < pcmData.length; i++) {
                int16Array[i] = Math.max(-32768, Math.min(32767, pcmData[i] * 32768));
            }
            if (this.ws.readyState === WebSocket.OPEN) {
                this.ws.send(int16Array.buffer);
            }
        };

        this.ws.onmessage = (event) => {
            const caption = JSON.parse(event.data);
            this.displayCaption(caption.text);
        };

        const source = audioContext.createMediaStreamSource(stream);
        source.connect(processor);
        processor.connect(audioContext.destination);
    }

    displayCaption(text) {
        // Rolling-window display (last 2-3 lines)
        const line = document.createElement('p');
        line.textContent = text;
        line.className = 'caption-line';
        this.captionDiv.appendChild(line);

        // Remove old lines
        while (this.captionDiv.children.length > 3) {
            this.captionDiv.removeChild(this.captionDiv.firstChild);
        }

        // Auto-scroll
        this.captionDiv.scrollTop = this.captionDiv.scrollHeight;
    }
}

How to Choose an STT Model for Captions?

Comparison between on-premise and cloud solutions:

For tasks requiring full data isolation (medical, government) — on-premise faster-whisper with Triton Inference Server. For typical broadcasts — cloud Deepgram or AssemblyAI. We will assess your project and suggest the optimal option. Request a preliminary audit — it is free and takes 30 minutes.

Display Requirements (WCAG 2.1)

/* Captions for hearing-impaired — WCAG 2.1 criterion 1.4.3 */
.caption-container {
    background-color: rgba(0, 0, 0, 0.85);
    color: #FFFFFF;
    font-size: 1.5rem;           /* minimum 24px */
    line-height: 1.6;
    padding: 12px 20px;
    border-radius: 4px;
    max-width: 80%;
    font-family: Arial, sans-serif;  /* high legibility */
}

/* High contrast (ratio 7:1 for AA+) */
.caption-line {
    color: #FFFFFF;
    text-shadow: 1px 1px 2px #000;
}

Integration with Zoom/Teams via Bot

# Zoom uses RTMP for streaming captions
import httpx

async def push_zoom_captions(meeting_id: str, caption_text: str, seq: int):
    """Send captions to Zoom via Closed Caption API"""
    async with httpx.AsyncClient() as client:
        await client.post(
            f"https://api.zoom.us/v2/meetings/{meeting_id}/live_streaming/captions",
            json={"text": caption_text, "seq": seq, "lang": "ru-RU"},
            headers={"Authorization": f"Bearer {ZOOM_JWT_TOKEN}"}
        )

What Is Streaming Transcription?

Streaming transcription is a technology where the model recognizes speech as audio fragments arrive, outputting partial results every 100–200ms. This allows captions to update smoothly, without pauses for the complete phrase. We use WebSocket to transmit audio fragments and receive text with timestamps.

Implementation Checklist

• [ ] Audit current audio channels and platform
• [ ] Choose STT model (on-premise/cloud)
• [ ] Calibrate for domain vocabulary
• [ ] Develop WebRTC/WebSocket server
• [ ] Integrate with Zoom, Teams, YouTube Live, RTMP
• [ ] Management interface and manual correction
• [ ] Documentation and operator training
• [ ] 6-month warranty

Estimated Timelines

Web captioning component — 1–2 weeks. Full platform integration — 2–3 weeks. To assess your project, contact us — describe the scenario and we will calculate the solution. Budget savings can reach 40% compared to manual captioning, with payback in 2–3 months. Get a consultation now!

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.