Voice Extractor

Voice Extractor

Extract authentic voice characteristics from transcripts to enhance author profiles.

Purpose

Transform transcript data (podcasts, interviews, meetings, videos) into actionable writing guidelines that capture an author's authentic voice, making AI-generated content sound natural and personal.

When to Use

• Author has recordings/transcripts of themselves speaking
• Want to capture authentic speaking patterns
• Need to enhance a manually-created author profile
• Building a new author profile from scratch using transcripts
• Refining an existing profile with more data

Workflow

1. Prepare Transcripts

Accept transcripts in these formats:

• Plain text: Speaker: text
• Timestamped: [00:01:23] Speaker: text or 59:54 Speaker: text
• Bracketed: [Speaker]: text
• WhatsApp: [17:30, 12/6/2025] Speaker: text
• SRT subtitles: Standard subtitle format

If user provides audio/video without transcript: Suggest transcription services:

• YouTube auto-captions (downloadable)
• Otter.ai, Descript
• OpenAI Whisper (local)
• Rev.com

2. Run Analysis

# List speakers in transcript
bun run "${CLAUDE_PLUGIN_ROOT}"/scripts/voice-extractor.ts --list-speakers transcript.txt

# Extract for specific speaker
bun run "${CLAUDE_PLUGIN_ROOT}"/scripts/voice-extractor.ts --speaker "Name" transcript.txt

# Multiple transcripts (more data = better profile)
bun run "${CLAUDE_PLUGIN_ROOT}"/scripts/voice-extractor.ts --speaker "Name" t1.txt t2.txt t3.txt

# Output JSON for direct use with author profiles
bun run "${CLAUDE_PLUGIN_ROOT}"/scripts/voice-extractor.ts --speaker "Name" --json transcript.txt

3. Enhance Author Profile

The extracted data enhances these author fields:

4. Manual Enhancement

After automated extraction, read transcript samples to identify:

• Unique phrases: Catchphrases missed by frequency analysis
• Humor style: Sarcasm, self-deprecation, wit patterns
• Story structure: How they set up anecdotes
• Cultural markers: Regional expressions, analogies
• Topic emphasis: What makes them animated

What Gets Extracted

Sentence Structure

• Average sentence length
• Variety (short/moderate/long/complex)
• Question frequency

Communication Style

• Enthusiasm (love, amazing, awesome)
• Hedging (maybe, perhaps, I think)
• Certainty (definitely, absolutely)
• Empathy (understand, appreciate)
• Directness (need to, must, bottom line)
• Storytelling (so, and then, eventually)
• Analytical (because, therefore, however)

Characteristic Expressions

• Fillers: "you know", "I mean", "like", "right"
• Hedgers: "kind of", "sort of", "I guess"
• Emphatics: "honestly", "literally", "definitely"

Vocabulary

• Signature words used frequently
• Vocabulary richness percentage

Integration with Author Profile

New Author from Transcript

# 1. Extract voice data
/article-writer:author analyze --speaker "John" transcripts/*.txt

# 2. Review and confirm extraction
# Claude will show extracted patterns

# 3. Add identity info
/article-writer:author add
# Answer: name, role, expertise, languages

# 4. Merge voice data
# Claude will combine extracted + manual data

Enhance Existing Author

# 1. Extract from new transcripts
/article-writer:author analyze --speaker "John" --author-id mwguerra new_podcast.txt

# 2. Review changes
# Claude shows what will be updated

# 3. Confirm merge
# Voice analysis data added to existing profile

Output Format

JSON Output (for merging)

{
  "voice_analysis": {
    "extracted_from": ["podcast_ep1.txt", "interview.txt"],
    "sample_count": 156,
    "total_words": 12450,
    "sentence_structure": {
      "avg_length": 14.5,
      "variety": "moderate length, conversational",
      "question_ratio": 12.3
    },
    "communication_style": [
      { "trait": "enthusiasm", "percentage": 28.5 },
      { "trait": "analytical", "percentage": 24.1 },
      { "trait": "directness", "percentage": 18.7 }
    ],
    "characteristic_expressions": [
      "you know",
      "I think",
      "the thing is",
      "at the end of the day"
    ],
    "sentence_starters": [
      "I think",
      "So the",
      "And then",
      "But the"
    ],
    "signature_vocabulary": [
      "actually",
      "basically",
      "approach",
      "strategy",
      "implementation"
    ],
    "analyzed_at": "2025-01-15T10:00:00Z"
  },
  "suggested_updates": {
    "tone": {
      "formality": 5,
      "opinionated": 7
    },
    "phrases": {
      "signature": ["you know", "the thing is", "at the end of the day"]
    },
    "vocabulary": {
      "use_freely": ["approach", "strategy", "implementation"]
    }
  }
}

Markdown Report

# Voice Analysis: John Smith

*Analyzed 156 speaking turns, 12,450 words*

## Speaking Style
- **Sentence length**: Moderate (~14 words avg)
- **Questions**: Uses questions occasionally (12%)
- **Vocabulary richness**: 45% unique words

## Communication Style
- **Primary**: Enthusiastic (28%)
- **Secondary**: Analytical (24%)
- **Tertiary**: Direct (19%)

## Characteristic Expressions
- "you know" (used 45x)
- "I think" (used 38x)
- "the thing is" (used 22x)

## Sentence Starters
- "I think..." (28x)
- "So the..." (19x)
- "And then..." (15x)

## Signature Vocabulary
**actually** (67x), **basically** (45x), **approach** (34x)

---

## Recommendations for Author Profile

Based on this analysis:
- Set formality to 5 (conversational but professional)
- Set opinionated to 7 (confident, uses "I think" but states opinions)
- Add signature phrases: "you know", "the thing is"
- Use vocabulary freely: approach, strategy, implementation

Quality Indicators

Good voice analysis needs:

• 100+ speaking turns for reliable patterns
• 5,000+ words for vocabulary analysis
• Multiple contexts (different topics/conversations)

Low data warning:

⚠️ Limited data: Only 23 speaking turns found.
   Results may not fully represent speaking patterns.
   Consider adding more transcripts.

References

• TRANSCRIPT-FORMAT-EXAMPLES.md
• VOICE-PROFILE-TEMPLATE.md

Transcript Format Examples

The extraction script supports multiple transcript formats. Here are examples of each:

Format 1: Simple Speaker Labels

John: This is what I said about the project.
Jane: And this is my response to that.
John: Let me follow up on that point.

Format 2: Bracketed Speakers

[John]: This is what I said about the project.
[Jane]: And this is my response to that.
[John]: Let me follow up on that point.

Format 3: Timestamped Transcripts (Bracketed)

[00:01:23] John: This is what I said about the project.
[00:01:45] Jane: And this is my response to that.
[00:02:10] John: Let me follow up on that point.

Format 4: Timestamped Transcripts (Inline)

59:54 Marcelo Guerra: Teve uma semana que eu quase fui bloqueado.
59:58 José Alves de Souza Neto: Mas você tava montando uso ou nem se preocupa?
01:00:14 Marcelo Guerra: Eu quase fui bloqueado, cheguei no limite.

Format 5: WhatsApp Export Format

[17:30, 12/6/2025] Fernando dos Santos: To meio que brincando e ajustando uns detalhes no código
[17:49, 12/6/2025] Marcelo Guerra: Ela também!
[17:49, 12/6/2025] Marcelo Guerra: Voltando agora
[17:52, 12/6/2025] +55 14 99166-6802: Tinha um código pra mexer mas é aniversário da sobrinha

Format 6: SRT Subtitle Format

Standard subtitle format from video editors, YouTube, etc. Note: SRT files don't have speaker labels, so all text is attributed to "Speaker".

1
00:00:00,560 --> 00:00:07,260
In this recording, I'm going to try and speak over the 30-second window.

2
00:00:07,260 --> 00:00:14,800
So I'm going to start with some random speech and have some pauses in between, like now.

3
00:00:14,800 --> 00:00:20,720
This will generate, as we know, another segment within the 30-second window that we're going
to start with.

Format 7: Uppercase Speaker Names

JOHN: This is what I said about the project.
JANE: And this is my response to that.
JOHN: Let me follow up on that point.

Speaker Name Variations

The script handles various speaker name formats:

• Simple names: John, Jane
• Full names: José Alves de Souza Neto, Marcelo Guerra
• Names with special characters: B&C - Fernando dos Santos
• Phone numbers (WhatsApp): +55 14 99166-6802
• Accented characters: João, José, André

Multi-line Turns

The script handles multi-line speaking turns where text continues without a new speaker label:

John: This is a longer thought that I'm expressing
and it continues on the next line without
a new speaker label appearing.
Jane: Got it, that makes sense.

Mixed Formats

The script can handle transcripts that mix formats, but for best results use consistent formatting throughout.

Common Sources

These formats are commonly exported from:

• YouTube auto-captions (SRT format)
• Otter.ai transcripts
• Descript exports
• Rev.com transcriptions
• WhatsApp chat exports
• Zoom transcripts
• Video editing software (SRT)
• Manual transcriptions

Voice Profile Template Reference

Standard Voice Profile Structure

A complete voice profile follows this structure:

# Voice Profile: [Name]

*Generated from X speaking turns, Y words analyzed*

## Overview
- **Speaking style**: [short and punchy/moderate length/longer, detailed/complex, elaborate]
- **Average sentence length**: [X] words
- **Question frequency**: [X]% of sentences are questions
- **Vocabulary richness**: [X]% unique words

## Communication Style
- **Primary tone**: [e.g., enthusiastic, analytical, empathetic] (X%)
- **Secondary characteristics**: [additional tones] (Y%)

## Characteristic Expressions
Phrases and fillers frequently used:
- "[phrase]" (used Xx)
- "[phrase]" (used Xx)

## Common Sentence Starters
- "[Starter]..." (Xx)
- "[Starter]..." (Xx)

## Signature Vocabulary
Words that appear frequently and characterize this speaker's vocabulary:
**word1** (Xx), **word2** (Xx), **word3** (Xx)

---

## Instructions for Writing in This Voice
1. **Sentence structure**: Use [style] sentences (avg ~X words)
2. **Engagement**: [Use questions frequently / Use statements more than questions]
3. **Tone**: Lean toward [primary tone] language
4. **Natural speech**: Occasionally include expressions like "[phrase]", "[phrase]"
5. **Vocabulary**: Favor words like: word1, word2, word3

Manual Analysis Checklist

When analyzing transcripts manually or supplementing automated analysis:

Speech Rhythm

• [ ] Fast/slow pacing
• [ ] Use of pauses (marked as "..." or explicit)
• [ ] Interruption patterns
• [ ] Topic transitions

Emotional Markers

• [ ] Enthusiasm indicators (exclamation marks, "love", "amazing")
• [ ] Hedging language ("maybe", "I think", "sort of")
• [ ] Certainty markers ("definitely", "absolutely")
• [ ] Humor patterns (sarcasm, self-deprecation, wit)

Unique Identifiers

• [ ] Catchphrases or signature expressions
• [ ] Cultural/regional expressions
• [ ] Professional jargon
• [ ] Metaphor preferences

Conversation Patterns

• [ ] How they greet/sign off
• [ ] How they handle disagreement
• [ ] How they show agreement
• [ ] How they ask questions