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.