AI-IVR Development: Intelligent Voice Menu

We often see customers losing patience while navigating 4–7 layers of touch-tone menus. "Press 1 for... press 2..." — such DTMF IVR irritates and filters out up to 40% of calls. Studies show 60% of users prefer a voice interface. Our approach is AI-IVR: the system understands free speech, determines intent in 1–2 questions, and routes the call without buttons. We have been implementing such solutions for the last 5 years, and customers report a 30–50% reduction in call handling time and a 20 percentage point increase in satisfaction. This is not just replacing buttons — it is a shift from a rigid decision tree to an adaptive dialogue based on LLM.

Problems AI-IVR Solves

Traditional DTMF menus struggle with non-standard requests, are difficult to update, and require memorizing sequences. AI-IVR based on NLP and ASR replaces the rigid decision tree with flexible dialogue. The customer says "I have a payment issue" — the system itself understands it needs to go to billing and transfers.

Comparison of DTMF-IVR and AI-IVR

Important: AI-IVR does not require full infrastructure replacement — we integrate it on top of your existing PBX via SIP or API.

How AI-IVR Understands Customer Intent

At its core is an LLM (e.g., GPT-4o or LLaMA 3) that receives the speech transcription (ASR) and determines intent. We use few-shot prompting: we pass a list of available destinations and example phrases in the prompt. The model returns a JSON with destination and confidence. If confidence is below 0.75, the system asks a clarifying question. This approach handles 95% of requests without involving a human agent.

class AIIVR:
    def __init__(self, routing_config: dict):
        self.destinations = routing_config["destinations"]
        self.llm = AsyncOpenAI()

    async def handle_call(self, call: IncomingCall) -> str:
        """Process incoming call — return destination"""
        # Greeting
        await call.say(
            "Hello! You have reached {company}. How can I help you?"
        )

        # Listen for intent (up to 8 seconds)
        user_input = await call.listen(timeout_sec=8, silence_threshold_ms=800)

        if not user_input:
            return await self.handle_silence(call)

        # Recognize intent and route
        route = await self.recognize_intent(user_input)

        if route["confidence"] >= 0.75:
            return await self.route_call(call, route["destination"])
        else:
            return await self.clarify_intent(call, user_input)

    async def recognize_intent(self, user_input: str) -> dict:
        destinations_description = "\n".join(
            f"- {d['id']}: {d['description']}"
            for d in self.destinations
        )

        response = await self.llm.chat.completions.create(
            model="gpt-4o-mini",
            messages=[{
                "role": "system",
                "content": f"""Determine where to route the call.\nAvailable destinations:\n{destinations_description}\nReturn JSON: {{"destination": "id", "confidence": 0.0-1.0}}"""
            }, {"role": "user", "content": user_input}],
            response_format={"type": "json_object"}
        )
        return json.loads(response.choices[0].message.content)

    async def route_call(self, call: IncomingCall, destination: str) -> str:
        dest = next(d for d in self.destinations if d["id"] == destination)

        # Confirm routing
        await call.say(dest.get("routing_message",
                                 f"Transferring you to {dest['name']}..."))

        if dest["type"] == "queue":
            await call.transfer_to_queue(dest["queue_id"])
        elif dest["type"] == "extension":
            await call.transfer(dest["extension"])
        elif dest["type"] == "bot":
            await call.transfer_to_bot(dest["bot_id"])

        return destination

Destination Configuration (YAML/JSON)

Routing paths are defined in a simple config — adding a new destination can be done without redeployment.

destinations:
  - id: technical_support
    name: "Technical Support"
    description: "Problems with service, errors, not working"
    type: queue
    queue_id: tech_support_q
    routing_message: "Connecting to technical support. Please wait."

  - id: billing
    name: "Billing and Invoices"
    description: "Payment issues, invoices, debts, tariffs"
    type: bot
    bot_id: billing_bot

  - id: sales
    name: "Sales and New Connections"
    description: "Subscribe to service, new contract, tariffs"
    type: queue
    queue_id: sales_q

What Business Metrics Does AI-IVR Improve?

According to Gartner, implementing intelligent IVR reduces load on first-line support by 40–60%. AI-IVR handles up to 80% of calls without agent involvement — 3 times more than DTMF. Average call time is cut in half. Savings on agents can reach 50% at scale. By reducing load, you can reassign staff to more complex tasks.

How AI-IVR Integrates with Your Existing PBX

Integration is done via SIP trunk or REST API. We connect to Asterisk, 1C-Bitrix24, Cisco, Genesys, and others without full infrastructure replacement. The system acts as an intermediary layer: it accepts the call, processes the dialogue, and hands control back to the PBX. For on-premise, we use vLLM with INT4 quantization — this reduces GPU costs by up to 4x compared to FP16.

Why Latency Is Critical for AI-IVR

If ASR + LLM takes more than 3 seconds, the user hangs up. We use streaming ASR (e.g., real-time Whisper) and intent caching for frequent requests. Typical p99 latency is 1.5–2 seconds. To reduce delays, we use model quantization and inference on Triton Inference Server.

Process of Work

We implement AI-IVR in several stages:

Total time for a typical project is 4–6 weeks. For a pilot with 3–5 destinations, 2–3 weeks.

What's Included

• Documentation: architecture diagram, prompt descriptions, instructions for adding destinations.
• Access: to logging and monitoring systems (Grafana, ELK), to API for self-updating configs.
• Training: a 4-hour workshop for the team on managing AI-IVR and fine-tuning new scenarios.
• Support: 2 weeks post-launch with daily standups and priority bug fixes.

Common Implementation Mistakes

1. Too many destinations in the prompt (more than 10) — reduces accuracy. Optimal is 5–7.
2. Lack of a fallback scenario for long silences — the system should re-ask or transfer to an agent.
3. Ignoring latency: if ASR + LLM takes more than 3 seconds, the user hangs up. We use streaming ASR and intent caching.
4. Poor handling of ambiguous requests — without a clarifying question, the customer may end up in the wrong place.

Request a demonstration of AI-IVR for your scenario — we will assess the project in one day and offer a solution with a result guarantee. Get a consultation from an engineer: our specialists will help you 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.