Therapy Session Transcriber

Part of the mindwork suite. Converts therapy session recordings into clean, formatted transcripts.

What It Does

1. Chunks large audio files at natural silence points (sentence boundaries)
2. Transcribes using OpenAI Whisper API
3. Formats as two-person conversation with Me: / Therapist: labels
4. Corrects grammar and transcription errors
5. Translates to English (for non-English sessions)

Prerequisites

• Docker installed and running
• OPENAI_API_KEY environment variable set
• The mindwork-transcribe Docker image built (see Setup)

Setup (One-Time)

Build the transcription Docker image from the plugin's transcribe directory:

# Find the mindwork plugin location and build the image
docker build -t mindwork-transcribe ~/src/mindwork/transcribe

Or if installed as a plugin, find the plugin path first:

# The transcribe tool is in the 'transcribe/' directory of this plugin
docker build -t mindwork-transcribe /path/to/mindwork/transcribe

Usage

Full Therapy Session Processing (Recommended)

Transcribe, format as conversation, and translate to English:

docker run --rm \
  -e OPENAI_API_KEY \
  -v $(pwd):/data \
  mindwork-transcribe /data/session.m4a --format-conversation --output /data/transcript.txt

Raw Transcription Only

Just transcribe without formatting or translation:

docker run --rm \
  -e OPENAI_API_KEY \
  -v $(pwd):/data \
  mindwork-transcribe /data/session.m4a --output /data/transcript.txt

With Speaker Diarization

For automatic speaker detection (alternative to --format-conversation):

docker run --rm \
  -e OPENAI_API_KEY \
  -v $(pwd):/data \
  mindwork-transcribe /data/session.m4a --diarize --output /data/transcript.txt

Only Chunk (No Transcription)

Split a large file into chunks for later processing:

docker run --rm \
  -v $(pwd):/data \
  mindwork-transcribe /data/session.m4a --no-transcribe --keep-chunks

Process Existing Chunks

Resume from previously created chunks:

docker run --rm \
  -e OPENAI_API_KEY \
  -v $(pwd):/data \
  mindwork-transcribe /data/chunks/ --format-conversation --output /data/transcript.txt

Options Reference

Option	Description
--output FILE	Save transcript to file (default: stdout)
--format-conversation	Format as Me/Therapist dialogue + translate to English
--diarize	Auto-detect speakers (uses gpt-4o-transcribe-diarize)
--no-transcribe	Only chunk, skip transcription
--keep-chunks	Preserve chunk files after processing
--model MODEL	whisper-1 (default, fast) or gpt-4o-transcribe (better accuracy)

Supported Audio Formats

mp3, mp4, m4a, wav, webm, ogg, flac

Configuration (mindwork.yaml)

If a mindwork.yaml config file exists, use it to determine output paths:

vault: ~/Therapy

sources:
  recordings:
    paths: [recordings/]

outputs:
  transcriptions: transcriptions/

Config locations (checked in order):

1. ./mindwork.yaml (current directory)
2. ~/.config/mindwork/config.yaml
3. ~/.mindwork.yaml

Default behavior (no config):

• Save to current directory or user-specified --output path

With config:

• Save to {vault}/{outputs.transcriptions}/{date}-{filename}.md
• Example: ~/Therapy/transcriptions/2024-01-15-session-001.md

See config/mindwork.example.yaml for full configuration options.

Output Format

With --format-conversation, output looks like:

**Me:** I've been feeling anxious about work lately. The deadlines keep piling up.

**Therapist:** That sounds overwhelming. Can you tell me more about what specifically triggers that anxiety?

**Me:** It's mostly when I have multiple projects due at the same time...

Cost Estimate

OpenAI Whisper API: ~$0.006/minute of audio GPT-4o for formatting/translation: ~$0.01-0.02 per session (varies by length)

A typical 50-minute session costs approximately $0.30-0.50 total.

Troubleshooting

"Docker image not found" Build the image from the plugin's transcribe directory:

docker build -t mindwork-transcribe /path/to/mindwork/transcribe

"OPENAI_API_KEY not set"

export OPENAI_API_KEY="sk-..."

"File not found" Ensure you're in the directory containing your audio file, or use absolute paths.

Transcription quality issues Try --model gpt-4o-transcribe for better accuracy (same price as whisper-1).