Real-time Streaming Speech Recognition (STT)

We have faced the challenge: a call center with 50 operators needing real-time subtitles for supervisors. Batch STT had a 5-second delay, missing important dialogue moments. For live conference subtitles, a 2-second delay is already unacceptable, and in voice assistant STT, every extra millisecond degrades UX. The solution is real-time streaming speech recognition with partial results over WebSocket. Over 5 years, we built an architecture that maintains 100–500 ms latency under any load.

What Problems Does Streaming Speech Recognition Solve?

• Latency: Without partial results, the user waits for the end of the phrase. Streaming outputs preliminary transcription every 200–400 ms. For a call center, this means instant reaction – the supervisor sees the text 200 ms after it is spoken.
• Pauses and overlaps: VAD endpointing correctly handles silence and overlapping speech. Setting aggressiveness=2 cuts 90% of pauses without loss of meaning.
• Real-time accuracy: Low-latency models (Deepgram Nova-2) achieve WER <5% even at 200 ms. The cost of Deepgram Nova-2 is $0.0043/min, 40% cheaper than Google STT. For 1000 hours per month, self-hosted faster-whisper on GPU costs about $60 (GPU rental), saving $2,520 per month compared to Deepgram.

Streaming Speech Recognition Implementation

A typical production architecture we have deployed:

Microphone → WebSocket (WSS) → FastAPI → STT Engine → NLP → Response

Key components are implemented in Python with async sockets, utilizing buffering optimization and endpointing algorithms.

WebSocket server on FastAPI

from fastapi import FastAPI, WebSocket
from faster_whisper import WhisperModel
import numpy as np
import asyncio

app = FastAPI()
model = WhisperModel("medium", device="cuda", compute_type="float16")

@app.websocket("/stream")
async def stream_stt(websocket: WebSocket):
    await websocket.accept()
    audio_buffer = bytearray()
    try:
        while True:
            chunk = await websocket.receive_bytes()
            audio_buffer.extend(chunk)
            if len(audio_buffer) >= 32000 * 2:  # 2 sec @ 16kHz 16-bit
                audio_array = np.frombuffer(audio_buffer, dtype=np.int16).astype(np.float32) / 32768.0
                segments, _ = model.transcribe(audio_array, language="ru")
                partial_text = " ".join([s.text for s in segments])
                await websocket.send_json({"type": "partial", "text": partial_text})
                audio_buffer = bytearray()
    except Exception:
        await websocket.close()

VAD (Voice Activity Detection)

We attach VAD before buffer accumulation: it cuts out silence, reducing the number of transcriptions.

import webrtcvad

vad = webrtcvad.Vad(2)
def is_speech(audio_chunk: bytes, sample_rate: int = 16000) -> bool:
    return vad.is_speech(audio_chunk, sample_rate)

For endpointing, we maintain a sliding silence window of 500–800 ms.

Client side

const socket = new WebSocket('wss://api.example.com/stream');
const mediaStream = await navigator.mediaDevices.getUserMedia({ audio: true });
const recorder = new MediaRecorder(mediaStream, { mimeType: 'audio/webm;codecs=opus' });
recorder.ondataavailable = (event) => {
    if (socket.readyState === WebSocket.OPEN) socket.send(event.data);
};
recorder.start(250); // 250ms chunks

How VAD Improves Streaming STT

Without VAD, the engine processes the entire audio stream, including silence. This increases token cost and latency. In practice, we saw p99 latency increase by 30% without VAD preprocessing. Additionally, proper VAD endpointing reduces computational overhead by 40%.

How to Select the Right STT Engine?

The choice between cloud and self-hosted depends on load, confidentiality requirements, and budget. According to official Deepgram documentation, Nova-2 has 180 ms latency at p95. To illustrate cost savings: for 1000 hours of audio per month, self-hosted faster-whisper on GPU costs approximately $60/month (GPU rental), while Deepgram would cost $2,580/month, saving $2,520/month. These cost figures demonstrate that streaming STT can be cost-effective, especially with self-hosted solutions.

Self-hosted solutions save up to 80% at volumes >1000 hours per month. For multilingual projects, Google and Azure are preferable due to broader coverage.

Ensuring STT Latency Under 400ms

Key factors: choosing a low-latency engine, optimizing VAD endpointing, and configuring buffering. For self-hosted, we use faster-whisper with CUDA and INT8 quantization – this reduces latency by 30% without accuracy loss. Plus pre-segmentation of audio via VAD to avoid transcribing silence.

Key Metrics to Monitor

• STT latency p99 – no more than 400 ms for self-hosted, 300 ms for cloud solutions.
• CPU/GPU utilization – to avoid overload under peak load.
• WER (Word Error Rate) – tracked on a sample set.
• Number of active sessions – important for auto-scaling.

Turnkey implementation process

1. Analysis: Define language, number of speakers, expected RPS, endpointing requirements.
2. Design: Build flow diagram, select engine, VAD, and dispatching method.
3. Development: Code WebSocket server, integrate STT, configure auto-scaling.
4. Testing: Generate synthetic RTP streams, measure p99 latency, memory leaks.
5. Deployment: Deploy in k8s with Helm, connect monitoring (Prometheus + Grafana).
6. Handover: Documentation, team training, codebase with comments.

Deliverables

• Architectural diagram and reasoning for choice
• Repository with Docker containers and Helm chart
• API documentation (OpenAPI)
• Integration with client SDKs (Web, iOS, Android – optional)
• Load test plan
• 1 month support

Timeline and cost

Cost is calculated individually per task. Get a project estimate – contact us.

Our experience

We have implemented streaming STT for 10+ projects, from call centers to live subtitles. Our experience includes integration with deep dialogue frameworks and configuration for high load (up to 1000 concurrent sessions). We guarantee p99 latency < 400 ms for self-hosted solutions on NVIDIA A10G. Certificated in CUDA (NVIDIA).

We are ready to implement streaming STT turnkey. Get in touch for a consultation – we will discuss your task and choose the optimal architecture.

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.