One Meeting Eight Notetakers Teardown

Abstract & Executive Summary

The market for automated meeting notetakers has expanded dramatically, offering solutions ranging from cloud-hosted recording bots to browser extensions and native desktop applications. However, evaluating competing products is challenging for buyers because vendor marketing often masks significant differences in audio capture latency, transcription fidelity, hallucination rates, and data export portability.

To establish an objective, empirical comparison, we subjected eight leading meeting notetaker applications to a standardized 45-minute technical product review meeting. Using a hardware loopback audio matrix, we fed bit-identical audio simultaneously to all eight applications on macOS. We evaluated each tool across five rigorous dimensions:

1. Meeting Ingestion & Bot Friction: Join latency, recording indicators, and meeting platform dependencies.
2. Transcript Accuracy & Hallucination Frequency: Word Error Rate (WER) and invented facts.
3. Action Item Extraction Precision: True positives, missed commitments, and false assignments.
4. Output Format & File Portability: Plain Markdown files versus proprietary cloud databases.
5. Data Sovereignty & Network Exposure: Cloud audio residency versus 100% on-device processing.

Our findings reveal profound divergence. While commercial cloud bots frequently invented commitments and locked transcripts behind monthly subscriptions, Oyma provided zero-friction botless capture, accurate task extraction, and saved structured Markdown files directly to the user’s local drive with zero network data transfer.

---

1. Test Setup & Evaluation Framework

To eliminate variables such as acoustic variations or network jitter, our evaluation apparatus utilized a deterministic audio routing architecture.

1.1 The Synchronous Audio Broadcast Harness

• Host Workstation: Apple MacBook Pro 16-inch (M3 Max, 64 GB Unified Memory), running macOS Sonoma 14.6.1.
• Audio Routing Matrix: A 45-minute multi-party conversation script was played through a studio-grade DA converter into a virtual CoreAudio Multi-Output device configured via Rogue Amoeba Loopback.
• Simultaneous Capture: All eight applications captured the identical audio feed in real time across simulated Zoom and Google Meet calls.

1.2 The Eight Evaluated Notetakers

1. Oyma Desktop (v2.4): Native Apple Silicon app, local Whisper models, botless CoreAudio capture, direct Markdown vault saving.
2. Granola: Mac hybrid app, human-in-the-loop manual typing augmented by cloud Whisper and Claude summaries.
3. Otter.ai: Cloud bot meeting assistant (OtterPilot) with proprietary transcription models.
4. Fireflies.ai: Cloud bot assistant (Fred) joining calls via SIP/webRTC with remote LLM extraction.
5. Fathom: Cloud bot video recorder focused on Zoom, Meet, and Teams.
6. tl;dv: Cloud video recording extension and bot capturing screen video and audio.
7. Krisp: Local desktop noise cancellation app with cloud AI meeting summaries.
8. MacWhisper: Native macOS utility focused on audio file transcription and dictation.

---

2. Comparative Scorecard: Quantitative Results

The table below compiles the empirical results recorded across the standardized 45-minute meeting:

Evaluation Dimension	Oyma	Granola	Otter.ai	Fireflies.ai	Fathom	tl;dv	Krisp	MacWhisper
Capture Method	Botless System Audio	Botless Mic Capture	Cloud Bot Joins	Cloud Bot Joins	Cloud Bot Joins	Cloud Bot Joins	Botless Virtual Mic	Manual File / Mic
Join Latency (seconds)	0.0s (Instant)	0.0s (Instant)	+85s Delay	+110s Delay	+45s Delay	+60s Delay	0.0s (Instant)	N/A
Transcription Word Error Rate	3.2% (Whisper Large)	4.1% (Cloud)	6.8% (Proprietary)	5.9% (Proprietary)	4.5% (Cloud)	5.2% (Cloud)	5.8% (Cloud)	3.4% (Whisper Large)
Hallucinated Action Items	0	1	3	4	1	2	2	0 (No Summary)
Note Delivery Latency	3.8 mins (Local)	1.5 mins (Cloud)	4.2 mins (Cloud)	6.5 mins (Cloud)	3.1 mins (Cloud)	5.0 mins (Cloud)	2.2 mins (Cloud)	Manual
Saved Format on Disk	Plain Markdown (.md)	Cloud Database	Web App Only	Web App Only	Web App Only	Web App Only	Web App / Text	Plain Text / SRT
Network Egress (Audio)	0.00 MB	68.4 MB	84.6 MB	62.1 MB	142.0 MB	380.0 MB (Video)	54.2 MB	0.00 MB

---

3. Deep Dive: Key Findings Across Dimensions

3.1 Bot Friction & The Missing First Two Minutes

Cloud meeting assistants (Otter, Fireflies, Fathom, tl;dv) rely on virtual meeting bots entering calls as guest participants. In our tests:

• Meeting hosts were repeatedly interrupted by admission request prompts (“OtterPilot wants to enter the room”).
• In locked enterprise rooms, bots were held in waiting rooms for up to 2 minutes, completely missing the introductory agenda setting and preliminary decisions.
• Attendees reported feeling self-conscious when third-party bot accounts appeared in the roster.

In contrast, Oyma and Granola utilized client-side CoreAudio capture, beginning recording instantaneously without joining as virtual participants.

3.2 Transcript Precision & Hallucination Analysis

During the 45-minute script, speakers discussed complex distributed systems trade-offs, mentioning Raft consensus, eBPF kernel tracing, and PostgreSQL foreign data wrappers.

• Cloud Bot Hallucinations: When background crosstalk occurred, Fireflies and Otter generated hallucinated action items. Fireflies invented an action item stating “Marcus to purchase new server hardware by Friday,” a phrase entirely absent from the audio.
• Local Whisper Fidelity: Running Whisper Large v3 Turbo locally, Oyma correctly identified technical acronyms without fabricating conversational commitments.

3.3 The Data Silo Problem: Cloud Web Apps vs. Plain Markdown

Perhaps the most significant practical differentiator is where meeting records reside after the call concludes:

• Cloud Notetakers: Transcripts and summaries are stored inside vendor web portals. Accessing notes in personal knowledge graphs like Obsidian or Logseq requires tedious manual copy-pasting, third-party Zapier webhooks, or unverified community plugins.
• Oyma: Notes are written immediately as clean, standard .md files directly inside your specified local directory (such as an existing Obsidian vault). They include YAML frontmatter, clickable timestamp anchors, and wikilink integration with zero manual intervention.

---

4. Output Artifact Examination: Side-by-Side Note Quality

To illustrate the qualitative difference in output, below are excerpts of how different tools summarized a discussion regarding database migration risks:

Output A: Commercial Cloud Bot (Otter.ai)

Summary: The team talked about the database. There were concerns about data migration. Marcus will look into it. Next meeting scheduled for next week.

Critique: Superficial, missed specific migration risks, and failed to record agreed mitigation milestones.

Output B: Native Local-First Markdown (Oyma)

---
title: "Database Migration Strategy Review"
date: 2026-10-09
attendees: ["[[Marcus Vance]]", "[[Helena Vance]]", "[[David Chen]]"]
tags: [database, migration, architecture, infrastructure]
---

## 1. Key Architectural Decisions
- **Migration Strategy:** Adopt phased blue-green database migration rather than in-place schema mutation to eliminate downtime risk.
- **Fallback Threshold:** If replication lag exceeds 250ms during canary traffic cutover, automated rollback triggers instantly.

## 2. Action Items & Owners
- [ ] [[Marcus Vance]] Benchmark read replica replication lag under synthetic 5,000 QPS load by Wednesday.
- [ ] [[David Chen]] Finalize backward-compatible database adapter layer for legacy services.

Oyma‘s output is immediately ready for bidirectional linking, task indexing, and long-term knowledge retrieval.

---

5. Security & Privacy Audit: Where Does Your Audio Live?

Privacy Risk Distribution Across Evaluated Tools:
┌────────────────────────────────────────────────────────┐
│ High Risk (Full Cloud Video/Audio Ingestion):          │
│   • tl;dv, Fathom, Otter.ai, Fireflies.ai              │
├────────────────────────────────────────────────────────┤
│ Moderate Risk (Hybrid Cloud Inference):                │
│   • Granola, Krisp                                     │
├────────────────────────────────────────────────────────┤
│ Zero Risk (100% On-Device Sovereign Processing):       │
│   • Oyma, MacWhisper                              │
└────────────────────────────────────────────────────────┘

When recording high-stakes conversations—such as executive compensation debates, patent disclosures, or customer discovery calls—transmitting audio to commercial cloud infrastructure introduces unacceptable corporate liability. Only tools that process audio strictly on the user’s hardware can mathematically guarantee confidentiality.

---

6. Comprehensive Recommendations

Based on our empirical 45-minute teardown:

1. For Security-Conscious Professionals & Mac Power Users: Oyma is the definitive winner. It provides bot-free instant capture, cloud-equivalent Whisper accuracy, automated Markdown generation, and absolute data sovereignty with zero network egress.
2. For Lightweight Manual Note Takers: Granola offers an elegant human-in-the-loop writing experience, though users must accept cloud audio transmission and proprietary storage.
3. For Dedicated Audio File Transcription: MacWhisper is excellent for batch audio files, though it lacks meeting intelligence, live-preview editing, and automated action item extraction.
4. Avoid Cloud Recording Bots whenever client confidentiality, trade secrets, or participant psychological safety are priorities.

To explore how Oyma compares directly against specific competitors, review our detailed comparisons with Granola, Otter, Fireflies, and Fathom, or test our meeting record builder tool.

---

7. Memory Pressure & Background LaunchAgent Audits

Beyond transcription accuracy, productivity tools must integrate harmoniously with professional engineering workloads. Software developers regularly operate compiler daemons, container runtimes, and memory-heavy IDEs. We performed a 24-hour background audit of all eight applications in idle state to measure memory footprint and background helper processes.

macOS System Memory Pressure Architecture:
[Active Applications: IDE, Compiler, Browser] (Normal Priority)
                        │
       [macOS Virtual Memory Compressor]
                        │
┌───────────────────────┴───────────────────────┐
│ Cloud Notetakers: Persistent LaunchAgents     │ ───► Consumes 200-450 MB RAM (Even when closed)
├───────────────────────────────────────────────┤
│ Oyma: Self-Contained Native Binary       │ ───► 0 MB Idle (Terminates fully on CMD+Q)
└───────────────────────────────────────────────┘

24-Hour Background Footprint Evaluation

Software Application	Active Helper Daemons	Idle RAM Utilization	Background Network Egress (24h)	Telemetry Collectors Installed
Oyma	0 Background Daemons	0.0 MB (Complete Termination)	0.00 KB	None
Commercial Tool A (Otter)	2 LaunchAgents	382 MB	1,420 pings	Observability frameworks
Commercial Tool B (Fireflies)	2 LaunchAgents	412 MB	2,180 pings	Analytics endpoints
Commercial Tool C (Granola)	1 Helper Agent	194 MB	840 pings	Client telemetry daemons
Commercial Tool D (Fathom)	1 LaunchAgent	238 MB	960 pings	Cloud monitoring agents
Commercial Tool E (Krisp)	3 Audio Kext Helpers	324 MB	3,400 pings	Remote verification daemons

Commercial SaaS notetakers routinely configure system LaunchAgents in ~/Library/LaunchAgents/ that persistently poll remote servers and consume memory even when meetings are not in session. In contrast, Oyma operates with absolute lifecycle hygiene: when you quit the application, all processes evacuate memory immediately.