AI Call Scoring System for Automated Operator Evaluation

Real case from our practice: our client — a retail chain with 300 calls per day. The QA team managed to listen to only 15% (45 calls). The missing 85% contained script violations, drops in empathy, and missed sales opportunities. Losses — up to 8% of revenue monthly (average 500k rub). We designed an AI system based on GPT-4o and Whisper that provided 100% coverage, increased scoring accuracy to 92% (compared to a human assessor), and reduced turnaround time by 70%. Savings amounted to over 400k rub monthly, with a payback period of 2–3 months.

Call scoring system (by definition Wikipedia) assigns a numeric score to each call based on a standardized methodology, creates operator rankings, and detects patterns requiring corrective training. The evaluation is not binary (ok/not ok) but multidimensional: greeting, hold, farewell, empathy, problem understanding, solution accuracy, GDPR compliance, AHT efficiency, and FCR.

What problems does AI scoring actually solve?

1. Selective control. A human assessor evaluates 10–20% of calls, the rest are a black box. AI processes 100%: every call gets a full scorecard. 2. Bias. Fatigue, subjective perception, different interpretation of criteria. LLM consistently applies the same methodology to all calls. 3. Lack of trends. Manual evaluation doesn't provide aggregate metrics. We build dashboards with moving averages, weekly/monthly trends, and heatmaps by violation type.

How we do it: stack and case

For evaluation, we use GPT-4o with response_format=json_object. The prompt contains all criteria with weights, an instruction "rate from 0 to 10" and a requirement to explain each rating. We apply fine-tuning (LoRA) to adapt to the specifics of operator speech. For transcription, we use Whisper with a target WER below 10%. In complex cases, we incorporate RAG with a knowledge base of scripts and frequent questions.

Example scorecard — Pydantic model:

from pydantic import BaseModel
from typing import Optional

class CallScorecard(BaseModel):
    call_id: str
    operator_id: str
    duration_seconds: float

    # Compliance
    greeting_score: float        # 0-10
    hold_procedure_score: float  # 0-10
    farewell_score: float        # 0-10
    gdpr_compliance: float       # 0-10

    # Quality
    problem_understanding: float # 0-10
    solution_accuracy: float     # 0-10
    empathy_score: float         # 0-10

    # Efficiency
    aht_relative: float          # 0-10 (relative to target AHT)
    first_call_resolution: float # 0 or 10

    # Sales/Upsell (if applicable)
    offer_made: Optional[float] = None
    offer_quality: Optional[float] = None

    @property
    def total_score(self) -> float:
        scores = [
            self.greeting_score * 0.10,
            self.hold_procedure_score * 0.05,
            self.farewell_score * 0.05,
            self.gdpr_compliance * 0.10,
            self.problem_understanding * 0.20,
            self.solution_accuracy * 0.25,
            self.empathy_score * 0.15,
            self.aht_relative * 0.05,
            self.first_call_resolution * 0.05,
        ]
        return round(sum(scores), 1)

Automatic scoring via LLM:

async def score_call_llm(transcript: dict) -> CallScorecard:
    full_dialog = format_dialog(transcript["turns"])

    response = await client.chat.completions.create(
        model="gpt-4o",
        messages=[{
            "role": "system",
            "content": """You are a service quality evaluation expert.
            Rate the call on each criterion from 0 to 10.
            Be objective, base your rating only on the text.
            Return JSON with fields: greeting_score, hold_procedure_score, farewell_score,
            gdpr_compliance, problem_understanding, solution_accuracy, empathy_score,
            first_call_resolution. For each field, add comment_FieldName with explanation."""
        }, {"role": "user", "content": full_dialog[:6000]}],
        response_format={"type": "json_object"}
    )

    data = json.loads(response.choices[0].message.content)
    return CallScorecard(
        call_id=transcript["call_id"],
        operator_id=transcript["operator_id"],
        duration_seconds=transcript["duration"],
        **{k: v for k, v in data.items() if not k.startswith("comment_")}
    )

To see how this works on your data, contact us—we'll send demo access.

How to set up a scorecard in 5 steps

1. Define call types (inbound/outbound, sales/support).
2. Select criteria from the library (15 base, custom can be added).
3. Assign weights to each criterion (sum to 1.0).
4. Set threshold values (e.g., AHT no more than 300 s).
5. Run a pilot on 100 calls and adjust the prompt based on results.

Comparison: AI vs human — who is more accurate?

Why does AI work more accurately than a human?

LLMs don't get tired, don't skip calls due to lack of time, and apply the same criteria to all dialogues. In our A/B test, AI showed 92% consistency with the control group (2 QA managers, 100 calls), while consistency between two assessors was 78%. This confirms that automated evaluation is not only faster but also more objective.

What's included in the work

• ETL architecture for call transcription (Whisper, Russian language recognition models).
• Integration with ACD/CRM — receiving calls and enriching with customer data.
• Scorecard of 15 criteria — weight customization for your scenario.
• LLM assistant — generating scorecard with explanations.
• Dashboard of ratings and trends — Grafana or React.
• Calibration and monitoring — automatic tracking of AI vs human correlation.

Process overview

1. Analytics and audit — review current evaluation criteria, scripts, call types.
2. Scorecard design — weights, AHT norms, FCR boundaries.
3. Integration and labeling — connect to telephony, gather historical recordings.
4. Training and calibration — run baseline, adjust prompt, fine-tune model if needed (LoRA).
5. A/B testing — compare AI with manual assessments.
6. Deployment and dashboards — deploy inference, connect alerts (Telegram, Slack).

Estimated timelines

Cost is calculated individually based on call volume, required latency, and need for fine-tuning. Request a consultation — we'll send a sample scorecard and estimate.

Calibration: how we maintain accuracy

Periodically we compare AI scores with manual QA manager scores. Target: Pearson correlation > 0.85. If correlation drops, we initiate recalibration on a fresh labeled dataset. The process includes:

• Collect new transcripts (100 calls)
• Labeling by two QA managers
• Training a LoRA adapter based on discrepancies
• A/B test the following week

Our team's experience — 5+ years in NLP and 30+ call center projects. We guarantee transparency: you get full access to prompts, weights, and evaluation logs. Request demo access to see results 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.