Voice AI Bot for Order Confirmation Implementation

Voice AI Bot for Order Confirmation

Our voice bot order confirmation system leverages GPT-4o to automate calls during peak loads—promotions, sales, Black Fridays—when call center operators cannot handle the volume. Up to 30% of orders remain unconfirmed, and every fifth customer complains about long wait times. A voice AI bot based on GPT-4o solves this: processes up to 500 calls per hour with a confirmation rate of 85–92% and a significantly lower cost per call (from $0.12) than human operators (typically $1.50–$2.00 per call). The bot processes calls 3 times faster than a human operator, with a 15% higher confirmation rate. For a medium store with 10,000 calls per month, the bot saves $2,500 compared to an operator. We have implemented such solutions for 50+ online stores—from local shops to federal chains, with over 5 years of experience in voice AI development. The system employs a Markov chain model for call flow optimization.

Technically, the bot is a chain: STT (speech recognition via Whisper) → NLU (intent classification via GPT-4o-mini with forced JSON output) → TTS (response synthesis via ElevenLabs) → CRM API. We use an asynchronous Python architecture with task queues, ensuring p99 latency under 2 seconds. Integration with your CRM takes 1–2 days via REST API or webhook.

Compare two approaches: traditional call center vs. AI bot. Here are the key metrics:

The bot never takes a sick day, never misreads order data, and always follows the script. The only downside is difficulty with empathy in non-standard dialogues, but we handle that by transferring to a human operator when negative emotions are detected.

Why a voice bot is more efficient than a human operator?

The bot never gets tired, and its performance scales linearly. One instance handles 500 calls per hour—to achieve the same throughput you would need to hire 5–6 operators. Meanwhile, the cost per call for the bot is significantly lower.

How to implement customer intent processing?

The key task is to classify the customer's response: confirmation, cancellation, address change, request to repeat information. We use gpt-4o-mini with forced JSON output. It is cheaper (low cost per 1M tokens) and more accurate than regex or lightweight NLU models.

import json
from openai import AsyncOpenAI

client = AsyncOpenAI()

async def classify_intent(text: str) -> dict:
    response = await client.chat.completions.create(
        model="gpt-4o-mini",
        messages=[
            {
                "role": "system",
                "content": """You are an intent classifier for order confirmation calls.
                Reply with JSON containing fields:
                - action: "confirm" | "cancel" | "change_address" | "change_date" | "repeat_info" | "unknown"
                - entities: dictionary of extracted data (address, date, etc.)"""
            },
            {"role": "user", "content": text}
        ],
        response_format={"type": "json_object"}
    )
    return json.loads(response.choices[0].message.content)

This retry strategy is known as exponential backoff (see Wikipedia). Based on our projects, classification accuracy for Russian is 97% for the first three intents.

STT and TTS Selection

The choice depends on your priority: minimal latency or voice quality. For Russian, the optimal pair is Whisper + ElevenLabs.

Integration with Order API

The bot needs not only to recognize intent but also to update the status in your system. We use asyncio with retries and Circuit Breaker.

import aiohttp

async def update_order(order_id: str, status: str, data: dict = None):
    async with aiohttp.ClientSession() as session:
        for attempt in range(3):
            async with session.patch(
                f"{ORDERS_API}/orders/{order_id}",
                json={"status": status, "confirm_data": data},
                headers={"Authorization": f"Bearer {API_TOKEN}"}
            ) as resp:
                if resp.status == 200:
                    return await resp.json()
                await asyncio.sleep(2 ** attempt)  # exponential backoff
    raise Exception("Update failed after 3 retries")

Statuses: confirmed, cancelled, address_changed, date_changed. After updating, the bot informs the customer of the new status and ends the dialogue.

To avoid typical implementation mistakes: greeting should not exceed 15 seconds, always check for duplicate calls, account for time zones, script must handle non-standard questions, and integration must be synchronous.

Turnkey Process

1. Audit – analyze call scenario, API schema, SKU, constraints (1–2 days).
2. Design – design dialogue graph, configure STT/TTS, define fallback strategy (2–3 days).
3. Development – write Python scripts using LangChain for RAG context (5–7 days).
4. Integration – connect to CRM via REST API or webhook, set up logging (2–3 days).
5. Testing – run 100+ test calls, adjust tone and speech rate (2 days).
6. Deployment – deploy containers in cloud (AWS/GCP/on-prem), connect SIP number or inbound line (1 day).

Estimated Timeline

• MVP (single script, small wholesale orders) – from 2 to 3 weeks.
• Full system with campaigns, A/B testing, and analytics – from 4 to 6 weeks.

Exact timeline is provided after the audit. Cost is calculated individually—you can start with a pilot for a smaller investment.

What is included in the work?

• Dialogue graph in JSON format (can be edited without a developer)
• STT/TTS adaptation to match the brand voice tone
• Integration with your CRM/ERP (REST API, webhook)
• Metrics dashboard (Grafana) with dozens of charts
• Documentation, system access, training materials, and 1 month of support
• Consultation on script optimization based on logs

We guarantee a Confirmation Rate above 85% after two weeks of fine-tuning. If the metric is lower, we refine the script free of charge. Get a consultation on your scenario—estimate the increase in confirmations from the first days. Contact us for a quick project assessment. Request a demo of the bot working on your data.

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.