Developing Voice AI Agents on VAPI: Implementation and Optimization

Is your client complaining that the bot interrupts or responds slowly? Most often the problem lies in incorrect interruption configuration and STT selection. We develop voice AI agents on VAPI — a platform that gives you full control over the stack: from transport to the model. Our experience: over 5 years and 50+ implemented projects, so we guarantee a reduction in p99 latency down to 800 ms and natural dialogue.

VAPI (Voice API) is an infrastructure platform for building voice AI agents with a developer focus. Unlike no-code solutions, VAPI provides full control over the stack: choice of STT provider (Deepgram, AssemblyAI), LLM (GPT-4o, Claude, Llama), TTS (ElevenLabs, Azure, OpenAI) and transport layer (WebRTC, PSTN, SIP). This allows creating agents with RAG, function calling, and custom voices that work 10 times faster than standard IVR systems.

VAPI Agent Architecture

Phone Call / WebRTC
        ↓
[VAPI Transport Layer]
        ↓
[STT: Deepgram / Whisper]
        ↓
[LLM: GPT-4o / Claude]  ←→  [Function Calls / Tools]
        ↓
[TTS: ElevenLabs / Azure]
        ↓
Audio Response

Why VAPI over Twilio or a Custom Solution?

Twilio Voice API is a low-level SIP stack where you have to connect each latency stage (STT, LLM, TTS) yourself. VAPI aggregates all stages in a single API call, handling timeouts and interruptions out of the box. The result: p99 latency 40% lower, and development costs 2–3 times less. For production, this means up to 40% savings on infrastructure and operators.

Creating an Agent via VAPI API

import requests
from typing import Optional

class VAPIAgentBuilder:
    def __init__(self, api_key: str):
        self.api_key = api_key
        self.base_url = "https://api.vapi.ai"
        self.headers = {
            "Authorization": f"Bearer {api_key}",
            "Content-Type": "application/json"
        }

    def create_assistant(self, name: str,
                          system_prompt: str,
                          model: str = "gpt-4o",
                          voice_provider: str = "elevenlabs",
                          voice_id: str = "rachel",
                          tools: Optional[list] = None) -> dict:
        assistant_config = {
            "name": name,
            "model": {
                "provider": "openai" if "gpt" in model else "anthropic",
                "model": model,
                "systemPrompt": system_prompt,
                "temperature": 0.7,
            },
            "voice": {
                "provider": voice_provider,
                "voiceId": voice_id,
                "speed": 1.0,
                "stability": 0.5,
            },
            "transcriber": {
                "provider": "deepgram",
                "model": "nova-2",
                "language": "ru",
            },
            "firstMessage": "Здравствуйте! Чем могу помочь?",
            "endCallMessage": "Спасибо за звонок. До свидания!",
            "endCallFunctionEnabled": True,
            "silenceTimeoutSeconds": 20,
            "maxDurationSeconds": 600,
        }

        if tools:
            assistant_config["model"]["tools"] = tools

        response = requests.post(
            f"{self.base_url}/assistant",
            json=assistant_config,
            headers=self.headers
        )
        return response.json()

    def create_tool(self, name: str,
                     description: str,
                     parameters: dict,
                     server_url: str) -> dict:
        return {
            "type": "function",
            "function": {
                "name": name,
                "description": description,
                "parameters": {
                    "type": "object",
                    "properties": parameters,
                    "required": list(parameters.keys())
                }
            },
            "server": {
                "url": server_url,
                "timeoutSeconds": 5,
            }
        }

    def create_outbound_call(self, assistant_id: str,
                              phone_number: str,
                              customer_data: dict = None) -> dict:
        payload = {
            "assistantId": assistant_id,
            "customer": {
                "number": phone_number,
                "name": customer_data.get("name", "") if customer_data else "",
            },
        }

        if customer_data:
            payload["assistantOverrides"] = {
                "variableValues": customer_data
            }

        response = requests.post(
            f"{self.base_url}/call",
            json=payload,
            headers=self.headers
        )
        return response.json()

    def setup_inbound_phone_number(self, phone_number: str,
                                    assistant_id: str) -> dict:
        payload = {
            "number": phone_number,
            "assistantId": assistant_id,
            "fallbackDestination": {
                "type": "number",
                "number": "+1234567890"
            }
        }

        response = requests.post(
            f"{self.base_url}/phone-number",
            json=payload,
            headers=self.headers
        )
        return response.json()

How to Reduce Latency to a Comfortable Minimum?

Latency consists of three stages: speech recognition (STT), model logic (LLM), and synthesis (TTS). In VAPI, you can influence each:

• STT choice: Deepgram Nova-2 gives ~250 ms with WER 8%, OpenAI Whisper ~600 ms but more accurate. For Russian-language projects, Whisper is often chosen.
• Interruptions: enabling interruptionsEnabled and setting numWordsToInterruptAssistant = 1 allows the user to interrupt the agent without delay.
• Transport: WebRTC is faster than PSTN — use it for clients in the region.
• Load balancing: deploy LLM on endpoints with low latency, for example via vLLM or Groq.

In practice, after optimization, p99 latency is 600–900 ms — a comfortable level for dialogue.

How to Configure Interruptions for Natural Dialogue?

VAPI allows fine-tuning parameters that affect conversational naturalness:

• interruptionsEnabled — allows the user to interrupt the agent. Critical for dialogue naturalness.
• backgroundDenoisingEnabled — background noise filtering via Krisp.
• numWordsToInterruptAssistant — how many words from the user are needed to interrupt the agent (recommended 1-2).
• backchannelingEnabled — agent utters “uh-huh”, “I see” during pauses.

{
  "model": {
    "provider": "openai",
    "model": "gpt-4o",
    "temperature": 0.7
  },
  "voice": {
    "provider": "elevenlabs",
    "voiceId": "rachel",
    "speed": 1.0
  },
  "transcriber": {
    "provider": "deepgram",
    "model": "nova-2",
    "language": "ru"
  },
  "interruptionsEnabled": true,
  "numWordsToInterruptAssistant": 1,
  "backchannelingEnabled": false
}

Integration with WebRTC for Web Calls

import Vapi from "@vapi-ai/web";

const vapi = new Vapi("YOUR_PUBLIC_KEY");

vapi.start({
  assistantId: "your-assistant-id",
});

vapi.on("call-start", () => console.log("Call started"));
vapi.on("call-end", () => console.log("Call ended"));
vapi.on("message", (message) => {
  if (message.type === "transcript") {
    console.log(message.role, message.transcript);
  }
  if (message.type === "function-call") {
    console.log("Tool:", message.functionCall.name);
  }
});

STT Provider Comparison in VAPI

Latency Optimization Parameters

What's Included in the Work

Each project includes:

• Agent architecture with optimal provider selection for your scenario.
• Implementation of Function Calls for integration with your systems (CRM, knowledge bases).
• Configuration of interruptions and limits for natural dialogue.
• Deployment to production (SageMaker, Vercel, own server).
• Code documentation and maintenance instructions.
• Warranty support for one month after launch.

Typical Mistakes When Developing a VAPI Agent

• Enabling interruptions without testing real scenarios: the agent does not listen to long responses.
• Using PSTN instead of WebRTC: latency is 1–2 seconds higher.
• Ignoring Function Call timeouts: if the endpoint takes longer than 5 seconds, the agent freezes.
• Missing fallback number: on error, the client should switch to an operator.

Timelines and Cost

A prototype of a voice agent with a basic scenario — from 2 to 3 days. A full-fledged solution with integrations, testing, and training — from 3 to 5 weeks. Cost is calculated individually for your project. Contact us to get a prototype in 2-3 days. Request a consultation to evaluate your scenario.

Source: VAPI REST API

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.