AI-Powered Meeting Minutes Automation

Manual meeting minutes consume 15–25% of participants' time. Companies lose up to 15% of their budget on ineffective meetings and manual note-taking. In a typical organization with 50 meetings per week, manual minutes take up to 20 person-hours, and up to 30% of tasks are lost or duplicated. Critical decisions get buried, deadlines slip, and accountability blurs. We solve this with a pipeline: audio recording → diarization → transcription → NLP extraction → structured minutes. It works in real time or post‑factum, integrates with Zoom, Google Meet, MS Teams, Slack, Jira, and Notion. Our solution achieves 2–3× better task extraction accuracy than off‑the‑shelf services thanks to custom NLP.

Problems Teams Face

• No single source of truth. Participants remember meetings differently. Decisions get re‑debated, tasks duplicated.
• Transcription without diarization. Simple services like Otter.ai don't separate speaker turns — the transcript is unreadable.
• Integration with task trackers. Even with a transcript, tasks must be entered into Jira/Notion by hand.

We close all three gaps. Our pipeline uses Whisper large‑v3 (Russian WER ~4.5% on clean recordings), pyannote.audio 3.1 for diarization (DER ~8% on multichannel conferences), and GPT‑4o for structure extraction. In sensitive scenarios we can deploy a local LLaMA 3 70B — confidential data never leaves the perimeter.

How We Customise AI Minutes for Your Infrastructure

A typical project takes 4–6 weeks and includes:

1. Audit current meetings – collect sample recordings, identify patterns (stand‑ups, code reviews, one‑on‑ones).
2. Choose the model – for short meetings (<1h) Whisper + GPT‑4o is enough; for long ones (3h+) we use VAD‑based chunking and parallel processing.
3. Set up integrations – via Zoom Recording API, Google Workspace Events, Microsoft Graph. Output is a webhook that triggers the pipeline.
4. Define the minutes format – Markdown for Confluence, custom templates for Notion, automatic task creation in Jira with deadlines from the transcript.

Real‑World Case

A fintech company with 200+ employees, weekly all‑hands for 150 people. Manual minutes took 8 person‑hours per week. We deployed the pipeline on their Kubernetes cluster with GPU T4. Results:

• Processing time for 1‑hour recording: 12 minutes (including diarization and NLP).
• Name recognition accuracy: 97% after fine‑tuning Whisper on corporate terms.
• Savings: 7 hours per week on minutes preparation alone.

Why Off‑the‑Shelf Solutions Fall Short

For startups with 5–10 meetings per week, off‑the‑shelf works. But for enterprises with confidential data, specialized terminology, and compliance requirements, our solution gives control and accuracy.

What's Included in the Delivery

• Transcription + diarization pipeline – Python code with CUDA support, unit‑tested.
• NLP module for extracting decisions and tasks – prompts tested on 500+ transcripts.
• Integrations – input: Zoom/Teams/Google Meet; output: Notion/Confluence/Jira/Slack.
• Documentation – README, architecture diagram, operations manual.
• Team training – 2‑hour workshop.
• 3‑month warranty – bug fixes, adaptation to API updates.

Get a free project assessment. Contact us — we'll show you how to cut minutes time by 5–10×. Consult with our AI engineer.

Implementation Details

Transcription with Diarization

import whisper
from pyannote.audio import Pipeline
import torch

class MeetingTranscriber:
    def __init__(self):
        self.whisper = whisper.load_model("large-v3", device="cuda")
        self.diarizer = Pipeline.from_pretrained(
            "pyannote/speaker-diarization-3.1",
            use_auth_token="HF_TOKEN"
        )

    def transcribe_with_speakers(self, audio_path: str) -> list[dict]:
        diarization = self.diarizer(audio_path)
        segments_by_speaker = [
            {"speaker": turn.speaker, "start": turn.start, "end": turn.end}
            for turn, _, _ in diarization.itertracks(yield_label=True)
        ]

        result = self.whisper.transcribe(audio_path, language="ru", word_timestamps=True)

        transcript = []
        for seg in result["segments"]:
            speaker = self._find_speaker(seg["start"], segments_by_speaker)
            transcript.append({
                "speaker": speaker,
                "start": seg["start"],
                "end": seg["end"],
                "text": seg["text"].strip()
            })

        return transcript

    def _find_speaker(self, timestamp: float, diar_segments: list) -> str:
        for s in diar_segments:
            if s["start"] <= timestamp <= s["end"]:
                return s["speaker"]
        return "UNKNOWN"

NLP Processing and Structure Extraction

from openai import AsyncOpenAI
import json

client = AsyncOpenAI()

async def extract_meeting_structure(transcript: list[dict]) -> dict:
    formatted = "\n".join([
        f"[{seg['speaker']} | {int(seg['start']//60):02d}:{int(seg['start']%60):02d}] {seg['text']}"
        for seg in transcript
    ])

    response = await client.chat.completions.create(
        model="gpt-4o",
        messages=[{
            "role": "system",
            "content": """Ты — ассистент для протоколирования встреч.
            Проанализируй транскрипт и верни JSON:
            {
              "summary": "краткое резюме 2-3 предложения",
              "participants": ["SPEAKER_00 = Иван Петров", ...],
              "agenda_items": [{"topic": "...", "discussion": "..."}],
              "decisions": [{"decision": "...", "context": "..."}],
              "action_items": [{"task": "...", "owner": "...", "deadline": "..."}],
              "next_meeting": "дата/условие следующей встречи если обсуждалась"
            }"""
        }, {
            "role": "user",
            "content": f"Транскрипт встречи:\n\n{formatted[:8000]}"
        }],
        response_format={"type": "json_object"}
    )

    return json.loads(response.choices[0].message.content)

Minutes Formatting and Export

def format_meeting_minutes(structure: dict, transcript: list[dict]) -> str:
    date = datetime.now().strftime("%d.%m.%Y")
    duration_min = int(transcript[-1]["end"] / 60) if transcript else 0

    md = f"""## Протокол встречи от {date}

**Продолжительность:** {duration_min} минут
**Участники:** {", ".join(structure.get("participants", []))}

### Краткое резюме

{structure.get("summary", "")}

### Принятые решения

"""
    for d in structure.get("decisions", []):
        md += f"- **{d['decision']}**\n  _{d.get('context', '')}_\n\n"

    md += "### Задачи\n\n"
    md += "| Задача | Ответственный | Срок |\n|--------|--------------|------|\n"
    for item in structure.get("action_items", []):
        md += f"| {item['task']} | {item.get('owner', '—')} | {item.get('deadline', '—')} |\n"

    return md

class MinutesExporter:
    async def to_notion(self, minutes: str, database_id: str): ...
    async def to_confluence(self, minutes: str, space_key: str): ...
    async def to_jira_tasks(self, action_items: list, project_key: str): ...
    async def to_slack(self, summary: str, channel_id: str): ...
    async def to_email(self, minutes: str, recipients: list[str]): ...

Webhook Integration (Zoom Example)

@app.post("/webhook/zoom/recording")
async def zoom_recording_webhook(payload: dict):
    if payload["event"] == "recording.completed":
        recording_url = payload["payload"]["object"]["recording_files"][0]["download_url"]
        meeting_id = payload["payload"]["object"]["uuid"]
        asyncio.create_task(process_meeting_recording(meeting_id, recording_url))
    return {"status": "ok"}

Timeline

Basic pipeline (transcription + NLP + Markdown) – 1–2 weeks. Full system with Zoom/Teams/Notion/Jira integrations – 4–6 weeks. Exact timeline depends on the number of recording sources and customisation requirements.

We have been in AI automation for over 5 years, delivering 30+ projects for finance, retail, and IT. We provide a warranty on pipeline functionality.

For a consultation and project assessment, contact us. Request a free audit of your meetings — we'll show you how much time you can save.

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.