Voice Analyser

Experimental library for extracting statistical voice models from your published writing. Generates immersive style guides that teach LLMs to replicate how you actually write - not through rules, but through examples and rhythm patterns.

Why This Exists

Traditional style guides list rules: "Use short sentences. Vary paragraph length. Include personal anecdotes."

This doesn't work. Writers don't follow rules - they channel voice.

This tool extracts the statistical fingerprint of your writing and presents it as immersive examples with annotations showing what makes each passage feel human. The goal is voice replication through pattern recognition, not rule compliance.

Status: Experimental. The approach works but is under active development.

Installation

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "voice-analysis": {
      "command": "npx",
      "args": ["-y", "@houtini/voice-analyser@latest"]
    }
  }
}

Config locations:

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

Restart Claude Desktop after saving.

Requirements

• Node.js 20+

Quick Start

1. Create an Output Directory

First, create a directory where your corpus and analysis will be stored:

C:\writing\voice-models\        (Windows)
~/writing/voice-models/          (Mac/Linux)

This directory will contain:

• Collected articles (markdown)
• Analysis JSON files
• Generated voice guides

2. Collect Your Writing

Collect corpus from https://yoursite.com/post-sitemap.xml 
Save as "my-voice" in "C:\writing\voice-models"

The tool needs:

• sitemap_url - Your XML sitemap URL
• output_name - A name for this corpus (e.g., "my-voice", "blog-posts")
• output_dir - The directory you created above

Example with all parameters:

Collect corpus from https://example.com/post-sitemap.xml
Output name: "technical-writing"  
Output directory: "C:\writing\voice-models"
Maximum articles: 50

3. Analyse Patterns

Analyse corpus "my-voice" in directory "C:\writing\voice-models"

This runs 14 analysers covering vocabulary, sentence structure, voice markers, argument flow, and paragraph transitions.

4. Generate Voice Guide

Generate narrative guide for "my-voice" in directory "C:\writing\voice-models"

Creates an immersive style guide with annotated examples at: C:\writing\voice-models\my-voice\writing_style_my-voice_narrative.md

Using the Voice Guide

Once generated, the voice guide can be loaded into any LLM conversation to help it write in your voice.

Loading the Guide

Load the file C:\writing\voice-models\my-voice\writing_style_my-voice_narrative.md 
and use it as a reference for all writing in this conversation.

Example Prompts for Content Generation

Blog post:

Using the voice guide as your reference, write a blog post about [topic].

Key requirements:
- Match the sentence rhythm patterns shown in the examples
- Use the conversational devices naturally (not forced)
- Include the micro-rhythms: mid-thought pivots, embedded uncertainty, present-tense immediacy
- Vary sentence length as shown in the statistical targets
- Use British/American spelling as indicated in the guide

Technical article:

Reference the voice guide and write a technical explanation of [concept].

Channel the voice by:
- Opening with the pattern shown in "Opening Moves" section
- Using specific product/tool names, not generic references  
- Including admissions of complexity or uncertainty where authentic
- Following the argument flow patterns from the guide
- Matching the punctuation habits (especially dash usage)

Product review:

Using the loaded voice guide, write a review of [product].

Capture the voice by:
- Starting with personal context (why you tested this)
- Blending technical specs with practical implications
- Using the transition patterns between paragraphs
- Including the "human tells" - parenthetical asides, mid-thought corrections
- Ending with the closing patterns shown in examples

Email/communication:

Write an email about [subject] using the voice patterns from the guide.

Focus on:
- Conversational markers appearing naturally
- Sentence length variation (some punchy, some complex)
- The hedging/confidence balance shown in statistics
- First-person usage matching the corpus frequency

Validation After Writing

The guide includes statistical targets. After writing, check:

Review what you just wrote against the voice guide metrics:
- Does sentence length variation match the target standard deviation?
- Is first-person frequency within the expected range?
- Are conversational markers present but not overused?
- Does the rhythm feel like the extended examples?

What Gets Analysed

Core Voice Patterns

• Vocabulary - Word choice, British/American markers, domain specificity
• Sentence structure - Length distribution, openers, complexity patterns
• Voice markers - First-person usage, hedging language, conversational markers
• Punctuation - Dash types, comma density, parenthetical frequency

Argument & Flow Patterns

• Argument flow - How you open, build, and close arguments
• Paragraph transitions - How ideas connect across paragraphs
• Conversational devices - "look", "frankly", "actually" and when they appear

Micro-Rhythm Detection

The guide annotates examples with invisible patterns that make writing feel human:

• Mid-thought pivots (comma before "and", "but", "so")
• Present-tense immediacy ("Right now, it's...")
• Embedded uncertainty ("I think", "probably")
• Casual sentence starters ("So,", "And,", "But,")
• Parenthetical asides
• Punchy fragments contrasting with longer sentences

Output Structure

your-output-directory/
└── corpus-name/
    ├── articles/                              # Collected markdown files
    ├── corpus.json                            # Metadata
    ├── analysis/                              # JSON analysis files
    │   ├── vocabulary.json
    │   ├── sentence.json
    │   ├── voice.json
    │   ├── paragraph.json
    │   ├── punctuation.json
    │   ├── function-words.json
    │   ├── argument-flow.json
    │   └── paragraph-transitions.json
    └── writing_style_[name]_narrative.md      # The voice guide

Minimum Corpus Size

• Minimum: 15,000 words (~20 articles)
• Recommended: 30,000 words
• Ideal: 50,000+ words

Below 15k words, statistical patterns become unreliable.

Tools Reference

collect_corpus

| Parameter | Required | Description | |-----------|----------|-------------| | sitemap_url | Yes | XML sitemap URL | | output_name | Yes | Corpus identifier (e.g., "my-voice") | | output_dir | Yes | Directory to store corpus | | max_articles | No | Limit (default: 100) | | article_pattern | No | Regex filter for URLs |

analyze_corpus

| Parameter | Required | Description | |-----------|----------|-------------| | corpus_name | Yes | Name from collect_corpus | | corpus_dir | Yes | Directory containing corpus | | analysis_type | No | full, quick, vocabulary, syntax |

generate_narrative_guide

| Parameter | Required | Description | |-----------|----------|-------------| | corpus_name | Yes | Name from analyze_corpus | | corpus_dir | Yes | Directory containing corpus |

Development

git clone https://github.com/houtini-ai/mcp-server-voice-analysis.git
cd mcp-server-voice-analysis
npm install
npm run build

Limitations

• Requires XML sitemap (RSS feeds not currently supported)
• Works best with consistent single-author content
• Mixed authorship or heavily edited content produces weaker signals
• The approach is experimental - results vary by writing style

Apache License 2.0 - Houtini.ai