AI Features

Custom AI Prompt Templates

Every meeting is different. A weekly standup needs a different summary format than a client discovery call or a board meeting. Karnyx ships with 7 built-in prompt templates that cover the most common scenarios, and you can create unlimited custom templates of your own.

Built-in Templates

Creating Custom Templates

Custom templates let you tailor AI output to your exact workflow. Navigate to Settings > AI > Prompt Templates > New Template to open the template editor.

Template Variables

Templates use variable injection with double-brace syntax. Karnyx replaces these variables with actual meeting data before sending the prompt to the AI model.

Example Custom Template

Here is a custom template for engineering incident retrospectives:

You are an incident retrospective analyst. Given the meeting transcript
and notes below, produce a structured retro document.

Meeting: {{meeting_title}}
Date: {{date}}
Duration: {{duration_minutes}} minutes
Attendees: {{attendees}}

TRANSCRIPT:
{{transcript}}

USER NOTES:
{{notes}}

HIGHLIGHTED SECTIONS:
{{highlights}}

Please produce:
1. **Incident Timeline** - chronological sequence of events
2. **Root Cause Analysis** - primary and contributing causes
3. **What Went Well** - things that helped during response
4. **What Went Wrong** - gaps in process or tooling
5. **Action Items** - assigned tasks with owners and deadlines
6. **Severity Assessment** - SEV level recommendation with justification

Preview Against Real Meetings

Before saving a template, you can test it against any previously captured meeting. Click "Preview" in the template editor, then select a meeting from the dropdown. Karnyx will run the template against that meeting's data and show you the AI output in real time. This lets you iterate on your prompt without waiting for a live meeting.

Multi-Language Transcription

Karnyx supports transcription in 28 languages with automatic language detection. When a meeting includes multiple languages, the system detects language switches mid-conversation and labels each segment appropriately.

Supported Languages

Auto-Detection

Language auto-detection is enabled by default. Karnyx analyzes the first 30 seconds of audio to determine the primary language, then continuously monitors for language switches. You can also pin a specific language in Settings > AI > Transcription Language if your meetings are always in the same language.

// Example transcript with language switches
[00:00:12] Sarah Chen (English):
  "Let's review the Q1 metrics before we switch to the Tokyo update."

[00:01:45] Takeshi Yamamoto (Japanese):
  "Q1の東京オフィスの数字について報告します。"

[00:03:22] Takeshi Yamamoto (English):
  "To summarize for everyone: Tokyo office exceeded targets by 12%."

[00:03:45] Sarah Chen (English):
  "Excellent. Let's move to the Paris update. Marie?"

[00:04:01] Marie Dubois (French):
  "Bien sûr. Les résultats du premier trimestre à Paris sont..."

// Karnyx labels each segment with the detected language
// and includes it in the transcript metadata

Voice Fingerprint Speaker ID

Karnyx creates a unique voice fingerprint for each person it encounters. After an initial enrollment, the system automatically identifies speakers in future meetings, labeling transcript segments with real names instead of generic "Speaker 1" / "Speaker 2" labels.

Enrollment

Voice fingerprint enrollment happens in two ways:

1. Automatic enrollment: After a meeting, Karnyx clusters speakers by voice similarity. When you label a speaker cluster with a real name (e.g., "Sarah Chen"), that voice is enrolled. The system needs at least 30 seconds of speech to create a reliable fingerprint.
2. Manual enrollment: Go to Settings > People > Voice Profilesand click "Enroll Voice". Record a 60-second sample or upload an existing audio clip of the person speaking.
3. Calendar-assisted enrollment: When attendees from your calendar speak in a meeting for the first time, Karnyx uses the attendee list as a hint and asks you to confirm the speaker identity after the meeting.

Automatic Matching

Once enrolled, voice fingerprints are matched automatically in every subsequent meeting. The matching engine runs in three stages:

1. Diarization: Audio is segmented into speaker turns using voice activity detection.
2. Embedding extraction: Each speaker turn is converted into a 256-dimensional voice embedding vector.
3. Fingerprint comparison: Embeddings are compared against the enrolled voice profile database using cosine similarity.

Confidence Thresholds

Karnyx uses a tiered confidence system to balance accuracy with coverage:

// Voice fingerprint matching result (internal representation)
{
  "speaker_turn_id": "turn_4a8f",
  "start_time": 124.5,
  "end_time": 138.2,
  "embedding": [0.0234, -0.1567, 0.0891, ...],  // 256-dim vector
  "matches": [
    { "profile": "Sarah Chen",    "score": 0.9451, "level": "high" },
    { "profile": "Lisa Wang",     "score": 0.7123, "level": "low" },
    { "profile": "Emily Parker",  "score": 0.5901, "level": "no_match" }
  ],
  "assigned_to": "Sarah Chen",
  "confidence": "high"
}

Highlight-Weighted Summarization

The notepad in Karnyx is not just for your own reference. What you type and highlight during a meeting directly influences how the AI generates its summary. This feature, called highlight-weighted summarization, ensures that the topics you care about most get the most attention in the output.

How It Works

1. Take notes during the meeting: Use the Karnyx notepad (always visible in the sidebar during a capture). Type freely, use bullet points, or shorthand.
2. Highlight important lines: Select any line and press Cmd+Shift+H to mark it as a highlight. Highlighted lines appear with an accent background.
3. AI weighting: When the meeting ends, Karnyx cross-references your highlights against the transcript. Transcript segments that correspond to highlighted notes receive a 3x weight boost in the summarization prompt.
4. Focused output: The resulting summary dedicates proportionally more space and detail to the topics you highlighted, while still covering the rest of the meeting.

Weight Levels

Example

Consider a 45-minute product sync where you highlighted "API rate limits discussion" in your notes. Even though that topic only occupied 5 minutes of the meeting, the summary will dedicate a full section to it, capturing the nuance of the rate limit proposal, objections raised, and the agreed-upon thresholds, while condensing the 20-minute roadmap discussion into a shorter overview.

// Internal weight calculation
{
  "meeting_id": "mtg_9x7k2",
  "segments": [
    {
      "topic": "Q1 Roadmap Review",
      "transcript_duration_pct": 44,
      "note_mentions": 2,
      "highlights": 0,
      "final_weight": 1.5,
      "summary_allocation": "20%"
    },
    {
      "topic": "API Rate Limits",
      "transcript_duration_pct": 11,
      "note_mentions": 1,
      "highlights": 1,
      "final_weight": 3.0,
      "summary_allocation": "35%"
    },
    {
      "topic": "Design System Updates",
      "transcript_duration_pct": 22,
      "note_mentions": 0,
      "highlights": 0,
      "final_weight": 1.0,
      "summary_allocation": "15%"
    },
    {
      "topic": "Hiring Pipeline",
      "transcript_duration_pct": 23,
      "note_mentions": 1,
      "highlights": 0,
      "final_weight": 1.5,
      "summary_allocation": "30%"
    }
  ]
}

Lenses

Lenses are reusable AI prompts that extract specific insights from your meetings. Unlike summaries which give you a general overview, Lenses are purpose-built to answer a specific question or extract a specific type of insight, such as competitor mentions, objection tracking, or budget analysis.

Karnyx ships with 25 built-in system Lenses organized across five categories (Sales, Product, Customer Success, Leadership, and Collaboration) and you can create unlimited custom Lenses. Run them against a single meeting or across multiple meetings at once using the / command menu.

Enhance Notes

The "Enhance Notes" feature takes your raw meeting notes and transforms them into polished, structured documents using AI. It cross-references your notes with the transcript to fill in gaps, correct details, and add context you may have missed while writing.

How It Works

1. After a meeting ends, open your notes from the meeting detail page.
2. Click the "Enhance" button in the notes toolbar.
3. Karnyx analyzes your notes alongside the transcript, identifies incomplete points, and produces an enhanced version.
4. Review the enhanced notes. Your original notes are preserved and you can toggle between the original and enhanced versions.

What Enhancement Does

• Gap filling: Adds details from the transcript that your notes missed
• Fact correction: Fixes names, numbers, and dates against what was actually said
• Structure improvement: Organizes bullet points into logical sections with headings
• Context addition: Adds speaker attribution and timestamp references
• Action item extraction: Pulls out action items mentioned in the transcript that you did not note

@Mentions

Use @mentions in your meeting notes to tag specific people. Type @ followed by a name to see an autocomplete dropdown of meeting participants and workspace members.

What @Mentions Enable

• Action item assignment: When you write "@Sarah to follow up on pricing", Karnyx automatically creates an action item assigned to Sarah.
• Notification delivery: Mentioned people receive a notification in the app (and via Slack/Teams if connected) with a link to the meeting and the context of the mention.
• People linking: @mentions create linkages in the participant graph, making it easier to search and cross-reference interactions.
• Summary emphasis: Topics associated with @mentioned people receive higher weighting in the AI summary.

Content Generation

Karnyx can generate polished content from your meeting data, saving you the time of writing it manually. Content generation is available from the meeting detail toolbar under the "Generate" dropdown.

Follow-up Emails

Generate a professional follow-up email based on the meeting outcomes. The email includes a summary of key points, action items with owners, and next steps. You can select the tone (formal, casual, or executive) and which recipients to address. The draft opens in your default email client for review before sending.

Blog Drafts

Turn brainstorming sessions and strategy meetings into blog post drafts. Karnyx extracts the key ideas, structures them into an article format with an introduction, main sections, and conclusion, and produces a first draft you can refine. Especially useful for content teams that want to capture thought leadership from meeting discussions.

Budget Extraction & Objection Tracking

Two specialized AI capabilities are built into the meeting processing pipeline for sales and customer-facing teams.

Budget Extraction

Karnyx automatically identifies and extracts financial figures, budget ranges, pricing discussions, and discount requests from your meeting transcripts. Extracted budget data appears in a dedicated section on the meeting detail page, including:

• Specific dollar amounts and ranges mentioned
• Budget constraints and approval thresholds
• Discount requests and negotiation points
• Payment terms and timeline expectations
• Speaker attribution for each financial mention

Objection Tracking

During sales and customer calls, Karnyx detects and catalogs objections raised by the other party. Each objection is tracked with:

• The objection statement and category (price, timing, competition, feature gap, etc.)
• Who raised the objection
• The response provided (if any)
• Whether the objection was resolved or left open
• Timestamp link to the exact moment in the transcript

AI Processing Pipeline

When a meeting ends, Karnyx runs a multi-stage AI pipeline to produce the final output. Understanding this pipeline helps you optimize your workflow and troubleshoot any issues.

1. Audio normalization: Volume leveling, noise reduction, and silence trimming (1-3 seconds).
2. Transcription: Speech-to-text using Whisper large-v3 locally on Apple Silicon, or cloud fallback for Intel Macs (30-90 seconds for a 1-hour meeting).
3. Speaker diarization: Voice activity detection and speaker clustering (10-20 seconds).
4. Voice fingerprint matching: Compares speaker embeddings against enrolled profiles (2-5 seconds).
5. Language detection: Per-segment language identification for multilingual meetings (1-2 seconds).
6. Note-transcript alignment: Cross-references your notes and highlights with transcript timestamps (1-2 seconds).
7. Summarization: Runs the selected prompt template with weighted inputs through the AI model (5-15 seconds).
8. Action item extraction: Identifies tasks, owners, and deadlines from the summary and transcript (2-5 seconds).