glebis/wispr-analytics

Wispr Analytics

Extract and analyze Wispr Flow dictation history from the local SQLite database. Combine quantitative metrics with LLM-powered qualitative analysis for self-reflection, work pattern recognition, and mental health awareness.

Data Source

Wispr Flow stores all dictations in SQLite at:

~/Library/Application Support/Wispr Flow/flow.sqlite

Key table: History with fields: formattedText, timestamp, app, numWords, duration, speechDuration, detectedLanguage, isArchived.

The user has ~8,500+ dictations since Feb 2025, bilingual (Russian/English), across apps: iTerm2, ChatGPT, Arc browser, Claude Desktop, Windsurf, Telegram, Obsidian, Perplexity.

Extraction Script

Run scripts/extract_wispr.py to pull data from the database:

# Get today's data as JSON with stats + text samples
python3 scripts/extract_wispr.py --period today --mode all --format json

# Get markdown stats for the last week
python3 scripts/extract_wispr.py --period week --format markdown

# Get text samples only for LLM analysis
python3 scripts/extract_wispr.py --period month --mode mental --texts-only

# Save to file
python3 scripts/extract_wispr.py --period week --format markdown --output /path/to/output.md

Period Options

• today -- current day (default)
• yesterday -- previous day
• week -- last 7 days
• month -- last 30 days
• YYYY-MM-DD -- specific date
• YYYY-MM-DD:YYYY-MM-DD -- date range

Mode Options

• all -- full analysis (default)
• technical -- filters to coding/AI tool dictations
• soft -- filters to communication/writing dictations
• trends -- focus on volume/frequency patterns
• mental -- all text, framed for wellbeing reflection
• prosody -- audio-based: pitch/intensity/voice-quality from recorded WAV (separate script scripts/extract_prosody.py; recent dictations only). See "Prosody Mode" below.

Comparison & Graphs

• --compare -- auto-compare with the equivalent previous period (week vs previous week, month vs previous month)
• --graphs PATH -- generate an HTML dashboard with Chart.js graphs (implies --compare). Graphs include: daily words overlay, hourly activity, category breakdown, top apps, language distribution

# Compare this month vs previous month (markdown)
python3 scripts/extract_wispr.py --period month --compare --format markdown

# Generate visual dashboard for week comparison
python3 scripts/extract_wispr.py --period week --compare --graphs /tmp/wispr-week.html

# Compare and save both markdown + graphs
python3 scripts/extract_wispr.py --period month --compare --format markdown --output report.md --graphs report.html

Prosody Mode (audio-based)

A standalone analysis mode -- a peer of technical/soft/trends/mental -- that analyzes how dictations sounded, not just what was said. It reads the recorded WAV audio stored in History.audio and uses Praat (via parselmouth) to extract prosodic features as gentle affect/energy proxies for self-reflection. Run it via the dedicated script scripts/extract_prosody.py.

Dependency

pip install praat-parselmouth

librosa/scipy/soundfile are acceptable fallbacks but the script uses parselmouth (Praat) as the gold standard.

Audio-retention caveat (read this first)

Wispr keeps the recorded audio only for recent dictations -- roughly the last ~900 of 16,000+ history rows. Older rows have their audio blob pruned after upload (and builtInAudio is always empty). So prosody is available for recent dictations only; for older periods the audio is gone and only timing-based metrics (rate, pauses) could ever be recovered. The script surfaces this honestly: every report opens with a coverage line (X of Y dictations in this period had retained audio) and logs when --limit truncates coverage.

What it measures

• Pitch (F0) via Praat to_pitch(), unvoiced frames ignored: mean, median, min, max, range, std, and CV (std/mean) as a monotone <-> expressive proxy.
• Intensity (dB): mean, range, std -- loudness dynamics.
• Voice quality: jitter (local), shimmer (local), and HNR (harmonics-to-noise ratio). Computed in try/except -- short/noisy clips that fail are skipped and counted (feature_failures).
• Tempo (from DB timing columns, not audio): speaking rate numWords / (speechDuration/60) WPM, and pause ratio (duration - speechDuration)/duration (clamped >= 0).
• By-language split (Russian vs English F0/rate differ -- kept separate so bilingual mixing doesn't muddy the signal) and a per-day trend table for multi-day periods.

Commands

# Prosody report for the last week (text)
python3 scripts/extract_prosody.py --period week

# Last month as JSON
python3 scripts/extract_prosody.py --period month --format json

# Specific day, raise the clip cap so coverage isn't truncated
python3 scripts/extract_prosody.py --period 2026-06-11 --limit 600

# Save to a file
python3 scripts/extract_prosody.py --period week --output /tmp/prosody-week.md

Args: --period (same semantics as extract_wispr.py: today/yesterday/week/month/YYYY-MM-DD/YYYY-MM-DD:YYYY-MM-DD), --format text|json, --limit N (cap clips processed, default 300 to bound runtime -- logs to stderr when it truncates), --output PATH. The DB is opened strictly read-only (mode=ro&immutable=1).

Performance: audio analysis is ~15-20s for 300 clips (slow vs SQL). The default --limit 300 keeps single runs fast; raise it for full coverage of a busy period.

Sanity expectations

Gleb is male, so expect mean F0 roughly 95-150 Hz (observed ~120 Hz). Russian typically shows slightly higher F0 and CV than English in the by-language split. F0 CV usually lands ~0.2-0.3.

How it ties into mental mode

Prosody complements the text-based mental mode with acoustic affect/energy proxies, framed the same way -- as reflection invitations, never diagnoses:

• F0 CV (pitch variability) as an engagement/expressiveness proxy: flatter = possibly tired/transactional, more varied = more animated.
• Speaking rate & pause ratio as energy / cognitive-load proxies.
• HNR drops can track vocal fatigue or strain.

Acoustic features are also shaped by microphone, room, a cold, and language -- so always name that uncertainty and compare like-with-like (same language, against the user's recent baseline). See the Prosody Mode template in references/analysis-prompts.md for the full interpretive prompt.

Workflow

Step 1: Extract Data

Run the extraction script with the requested period and mode. Use --format json for full data or --texts-only for LLM analysis focus.

Step 2: Present Quantitative Stats

Display the quantitative summary first:

• Total dictations, words, speech time
• Category breakdown (coding, ai_tools, communication, writing, other)
• Language distribution
• Hourly activity pattern
• Daily trends (for multi-day periods)
• Top apps

Step 3: Perform Qualitative Analysis

Read references/analysis-prompts.md to load the appropriate analysis template for the requested mode. Then analyze the text samples using that template.

For each mode:

Technical: Focus on what was worked on, technical decisions, context-switching patterns, productivity assessment.

Soft: Focus on communication style shifts, language-switching patterns, audience adaptation, interpersonal dynamics.

Trends: Focus on volume changes, time-of-day shifts, app migration, behavioral change hypotheses.

Mental: Focus on energy proxies, sentiment signals, rumination detection, activity pattern changes. Frame all observations as invitations for self-reflection, never as diagnoses. Use language like "you might notice..." or "this pattern could suggest..."

All: Combine all four perspectives into a unified reflection.

Step 4: Output

Default output location: meta/wispr-analytics/YYYYMMDD-period-mode.md in the vault.

File format:

---
created_date: '[[YYYYMMDD]]'
type: wispr-analytics
period: [period description]
mode: [mode]
---

# Wispr Flow Analytics: [period]

## Quantitative Summary
[stats from Step 2]

## Analysis
[qualitative analysis from Step 3]

## Reflection Prompts
[3-5 questions based on observations]

If the user requests console-only output, skip file creation and display directly.

App Category Mapping

The extraction script categorizes apps:

• coding: iTerm2, cmuxterm, VS Code, Windsurf, Zed, Cursor, Terminal
• ai_tools: ChatGPT, Claude Desktop, Perplexity, OpenAI Atlas, Codex
• communication: Telegram, Messages, Slack, Zoom
• writing: Obsidian, Notes, Chrome, Arc browser

Dictionary Management

Manage Wispr Flow's dictionary for better recognition accuracy. The dictionary JSON is version-controlled in ~/ai_projects/claude-skills/wispr-analytics/data/dictionary.json.

Dictionary Script

Run scripts/wispr_dictionary.py for all dictionary operations:

# Check database health and dictionary stats
python3 scripts/wispr_dictionary.py check

# List all entries (safe while Wispr is running)
python3 scripts/wispr_dictionary.py list
python3 scripts/wispr_dictionary.py list --filter "claude"

# Export dictionary to JSON (safe while running)
python3 scripts/wispr_dictionary.py export

# Suggest new entries by analyzing ASR vs formatted text differences
python3 scripts/wispr_dictionary.py suggest --days 30 --min-freq 3

# Propose snippets + replacement rules + vocab from dictation logs (safe while running)
python3 scripts/wispr_dictionary.py propose --days 30 --min-freq 3
python3 scripts/wispr_dictionary.py propose --days 90 --min-freq 2 --format json

# Add a single term (requires Wispr Flow to be QUIT)
python3 scripts/wispr_dictionary.py add "Gastown"
python3 scripts/wispr_dictionary.py add "cloud code" "Claude Code"

# Remove an entry (requires Wispr Flow to be QUIT)
python3 scripts/wispr_dictionary.py remove "old term"

# Import from JSON (requires Wispr Flow to be QUIT)
python3 scripts/wispr_dictionary.py import --dry-run
python3 scripts/wispr_dictionary.py import

Dictionary Safety Rules

CRITICAL: Wispr Flow must be quit before any write operations (add, remove, import). The script enforces this automatically. Read operations (export, list, suggest, check) are safe while Wispr is running.

Writing to the SQLite database while Wispr Flow has it open causes index corruption. Always:

1. Check if Wispr is running: pgrep -f "Wispr Flow"
2. If running, ask user to quit first (Cmd+Q)
3. After writes, run check to verify integrity
4. Restart Wispr Flow

Dictionary Entry Types

• Recognition terms (phrase only): teaches Wispr to hear the word correctly (e.g., "Gastown", "LLM", "subagent")
• Replacement rules (phrase → replacement): auto-corrects mishears (e.g., "cloud code" → "Claude Code", "клод дизайн" → "Claude Design")
• Snippets (isSnippet=true): text expansion shortcuts (e.g., "my email" → "[email protected]")

Propose Replacements & Snippets

suggest only catches ASR mishears. propose is the broader, human-style review: it reads recent dictation logs and proposes dictionary additions in three categories, skipping anything already in the dictionary. It is read-only and safe while Wispr Flow is running -- it never writes to the database.

# Default: last 30 days, terms seen >= 3 times
python3 scripts/wispr_dictionary.py propose

# Wider net, machine-readable
python3 scripts/wispr_dictionary.py propose --days 90 --min-freq 2 --format json

Flags: --days (history window), --min-freq (minimum occurrences), --format (text default, or json).

The three categories:

1. Snippet candidates (highest leverage, most underused): recurring URLs, emails, and phone numbers, plus repeated boilerplate sentences/intros/sign-offs (>= 8 words, counted by normalized verbatim frequency). Each proposal includes a short My X trigger phrase + the full expansion.
2. Replacement-rule candidates: recurring ASR mishears (shares code with suggest via the find_mishears helper).
3. Vocab candidates: frequently-dictated capitalized/technical terms (e.g. HTML, LinkedIn, SDK) that may be mis-recognized -- teach Wispr the spelling.

Each proposal prints a frequency count and a ready-to-run add command line.

Snippets are the single highest-leverage, most underused dictionary feature. A user with 1,600+ dictations/month often has only a handful of snippets. One My GitHub -> URL snippet saves dictating (and mis-dictating) a URL dozens of times. Always foreground snippet candidates first.

Suggested workflow

1. Run analytics (extract_wispr.py) to understand volume and where snippets pay off.
2. Run propose (safe while Wispr runs).
3. Present the grouped proposals to the user -- snippets first, then replacement rules, then vocab. Frame snippets as the big win.
4. After the user approves and quits Wispr Flow (Cmd+Q), run the approved add lines (snippets need add "My X" "expansion").
5. Run python3 scripts/wispr_dictionary.py check to verify integrity.
6. Restart Wispr Flow.

Proactive Dictionary Improvement Workflow

When running analytics, also check for dictionary improvement opportunities:

1. Run propose to surface snippets, replacement rules, and vocab in one pass (or suggest for mishears only)
2. Compare asrText vs formattedText for patterns
3. Look for Russian/English code-switching mishears
4. Check for new technical terms the user started using
5. Export updated dictionary and commit to git

Notes

• For analytics: the database is read-only; analytics never modifies Wispr data
• For dictionary: writes require Wispr Flow to be quit first
• Text samples are capped at 100 per extraction to manage context window
• For multi-day periods, daily trend tables help visualize changes
• Bilingual dictations are common; analysis should honor both Russian and English
• The asrText field contains raw speech recognition before formatting -- useful for detecting speech patterns vs formatted output
• Dictionary JSON is stored at ~/ai_projects/claude-skills/wispr-analytics/data/dictionary.json for version control