TalkTrack

Record, transcribe, and identify speakers from your calls — all locally on your machine. Free and open-source alternative to Evaer, Otter.ai, and Fireflies.

TalkTrack is a Windows desktop app for recording and transcribing Microsoft Teams calls, Zoom meetings, Google Meet sessions, and any other audio app. It uses Faster Whisper for local speech-to-text and pyannote.audio for speaker identification. Everything runs offline — no cloud services, no subscriptions, no data leaves your PC.

Why TalkTrack?

• No cloud uploads — your meeting recordings and transcripts stay on your machine
• No subscriptions — free and open-source, no monthly fees like Otter.ai or Fireflies
• Works with any app — Microsoft Teams, Zoom, Google Meet, Discord, Slack huddles, WebEx, or any app that plays audio
• Per-app capture — on Windows 11, record only your call app without picking up Spotify or YouTube in the background
• AI-powered — state-of-the-art Whisper speech recognition + pyannote speaker diarization, running locally on your hardware

Features

• Record calls with Record / Pause / Resume / Stop controls, live timer, and level meters
• Per-app audio capture (Windows 11) — pick specific apps like Teams or Chrome
• System audio capture (Windows 10+) — WASAPI loopback for all system audio
• Dual-channel recording — microphone + system/app audio captured separately
• Auto-stop recording — detects when your call app goes inactive and offers to stop
• Auto-start recording — optionally start recording when a selected app joins a call (Settings > General)
• Local transcription — Faster Whisper (OpenAI Whisper), no internet required
• Speaker diarization — two modes:
  • Simple (no setup): labels "You" vs "Remote" from mic vs system channels
  • Full (pyannote.audio): identifies individual speakers with a free HuggingFace token
• AI assistant — optional AI-powered meeting summaries, action items, and transcript chat (supports Claude, OpenAI, Grok, Gemini, Mistral, Groq, or local models)
• Per-provider AI settings — API keys and models stored separately per provider, switch without losing config
• Manual AI generation — generate or regenerate summaries and action items on demand
• Notes in AI context — call notes are included in AI summary and action item generation
• Interactive transcript — click any segment to replay its audio, edit text inline, assign speaker names
• Play All — sequential playback of the full transcript with real-time line highlighting and auto-scroll
• Export to TXT, SRT (subtitles), or JSON
• Call notes with timestamp insertion
• Recording browser — browse, replay, and bulk-delete past recordings (multi-select with Ctrl/Shift+click)
• Hidden devices filter — hide unwanted audio devices (e.g., Voicemeeter) from dropdowns via Settings
• Remembers capture settings — capture mode and selected apps persist across sessions
• Min recording length — automatically discard recordings shorter than a configurable threshold (Settings > General)
• Custom app icon — first-run Start Menu shortcut (targeting the venv) gives the correct Windows taskbar icon
• Collapsible audio sources — compact UI with expandable source selector
• GPU/CUDA detection — System Status panel detects your GPU and guides CUDA setup
• File logging — all errors logged to ~/.talktrack/talktrack.log with crash dialog
• Bug reporting — submit issues directly from the app via pre-filled GitHub issues
• Dark theme UI (Catppuccin Mocha palette)
• Guided setup wizard for HuggingFace / pyannote configuration
• Auto-install dependencies — first launch detects and installs required packages

Quick Start

Prerequisites

• Windows 10 or 11
• Python 3.10+
• A microphone

Install & Run

git clone https://github.com/ObscureAintSecure/TalkTrack.git
cd TalkTrack

Just double-click start.bat. On first launch it sets up an isolated environment and installs dependencies automatically — no manual steps needed.

Note: dependencies are installed into a project-local virtual environment (.venv), not your global Python. This keeps heavy packages like PyTorch and pyannote.audio from polluting or upgrading packages in your system Python.

Recommended: uv

If uv is installed, start.bat uses it automatically. You can also drive it directly:

uv sync            # create .venv and install pinned dependencies from uv.lock
uv run python main.py

uv sync is reproducible (it installs the exact versions in uv.lock) and fast. By default it installs CPU PyTorch, which works on any machine.

GPU acceleration (NVIDIA, optional)

If you have an NVIDIA GPU, install the CUDA build of PyTorch for much faster transcription and diarization. CPU and GPU builds are mutually exclusive extras — pick one:

uv sync --extra cuda     # CUDA 12.6 build (NVIDIA GPU)
# CPU is the default; uv sync --extra cpu pins the CPU build explicitly

Without uv:

pip install torch torchaudio --index-url https://download.pytorch.org/whl/cu126

Confirm it worked in Help > System Status (GPU Acceleration should read "detected with CUDA 12.6"), then set Compute Device to CUDA in Settings.

Windows note: if uv sync --extra cuda fails with failed to rename ... Access is denied (os error 5), antivirus is locking the large download in uv's cache. Retry, run uv cache clean first, or add %LOCALAPPDATA%\uv\cache to your antivirus exclusions. The plain-pip command above sidesteps it.

Without uv

If uv isn't installed, start.bat falls back to a local .venv created with Python's built-in venv + pip. To do it manually:

python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
python main.py

requirements.txt is kept in sync with pyproject.toml for users who prefer plain pip.

Custom taskbar icon

On first run TalkTrack offers to add a Start Menu shortcut (also available any time via Help > Add to Start Menu). The shortcut targets the project's .venv interpreter and carries the app icon, so Windows shows the correct taskbar icon. Launch from that shortcut, then right-click the running taskbar icon and choose Pin to taskbar if you want it pinned.

For troubleshooting, use start_debug.bat which shows a console window with log output.

Speaker Diarization (Optional)

For multi-speaker identification (speaker 1, speaker 2, etc...), TalkTrack uses pyannote.audio which requires a free HuggingFace account. On first launch, a setup wizard walks you through the steps:

1. Create a free account at huggingface.co
2. Accept the model license at pyannote/speaker-diarization-community-1
3. Create an access token at huggingface.co/settings/tokens
4. Paste the token into the wizard (or Settings > Transcription)

Without this setup, TalkTrack still works — it just labels speakers as "You" and "Remote" based on audio channels.

Usage

1. Start your call in Teams, Zoom, or any app
2. In TalkTrack, select your microphone and the app to capture
3. Click Record (or enable auto-record in Settings > General to start automatically when your app joins a call)
4. When done, click Stop — transcription starts automatically
5. Review the transcript, assign speaker names, use Play All to follow along, then export

Windows 11: Select specific apps (Teams, Chrome, etc.) in the app picker to capture only their audio.

Windows 10: Captures all system audio via WASAPI loopback.

How Transcription Works

TalkTrack uses Faster Whisper, a CTranslate2-optimized version of OpenAI's Whisper model. Everything runs locally — no audio is sent to any server.

Pipeline

1. Recording — Audio is captured as dual tracks: your microphone and system/app audio. Both are saved as WAV files alongside a combined_audio.wav used for transcription.
2. Transcription — When recording stops, Faster Whisper processes combined_audio.wav and produces timestamped text segments. VAD (Voice Activity Detection) filtering skips silence automatically.
3. Speaker Diarization — Speakers are identified and labeled on each segment (see below).
4. Review — The interactive transcript viewer lets you play back individual segments, edit text inline, and assign friendly names to speakers. Changes are saved to transcript.json and speaker_names.json in the recording directory.

Whisper Models

Choose a model in Settings > Transcription based on your speed/accuracy needs:

Models are downloaded automatically on first use and cached locally. No internet is needed after the initial download.

CPU vs GPU

• CPU (int8 quantization) — works on any machine, no extra setup. Good enough for most use cases.
• CUDA (float16) — significantly faster if you have an NVIDIA GPU with CUDA installed. Select "CUDA (NVIDIA GPU)" in Settings > Transcription > Compute Device.

TalkTrack automatically detects your GPU and CUDA availability in the System Status panel (Help > System Status) and shows guidance if CUDA isn't properly configured.

Language

By default, Whisper auto-detects the spoken language. You can set a specific language in Settings > Transcription > Language (e.g., en, es, de) to improve accuracy and speed if you know what language will be spoken.

Speaker Diarization

Speaker diarization identifies who is speaking at each point in the transcript. TalkTrack offers two modes:

Simple Mode (No Setup)

Works out of the box. Compares audio energy between your microphone track and the system/app audio track to label each segment as "You" or "Remote". Best for 1-on-1 calls.

Full Diarization (pyannote.audio)

Uses the pyannote.audio neural pipeline to identify individual speakers (SPEAKER_00, SPEAKER_01, etc.). Works for any number of participants. Requires a free HuggingFace account — see Speaker Diarization setup above.

You can optionally set min/max speaker counts in Settings to help the model when you know how many people are on the call.

After diarization, use the Speaker Name Panel in the transcript viewer to map generic labels (SPEAKER_00) to real names. Names are saved per recording and included in exports.

AI Assistant (Optional)

TalkTrack integrates with AI providers to generate meeting summaries, extract action items, and chat with your transcripts. AI features are entirely optional — the core recording and transcription works without them.

Supported Providers

SDK packages are installed automatically when you select a provider — no need to install them manually. Configure your provider and API key in Settings > AI.

To pre-install a provider (e.g. for offline setup), use the optional extras:

uv sync --extra claude       # or: openai, grok, gemini, mistral, groq, local, all-ai
# without uv:  pip install ".[claude]"

Export Formats

All exports include speaker names if assigned.

Settings

Access via the gear icon or Edit > Settings:

Project Structure

TalkTrack/
  main.py                    # Entry point, logging, crash handling
  start.bat                  # Launcher (uv-first, .venv isolation, pip fallback)
  start_debug.bat            # Debug launcher with console output
  requirements.txt           # Dependencies
  resources/
    style.qss                # Dark theme stylesheet
    talktrack.ico             # App icon (multi-size)
    arrow_up.png, arrow_down.png  # QSpinBox arrow icons
    build_ico.py              # Rebuild .ico from source PNGs
    TT_icon_*.png             # Icon source files (32-512px)
    TT_logo_*.png             # Logo files for branding
  app/
    main_window.py           # Main window + orchestration
    audio/
      segment_player.py      # Audio clip playback
    recording/
      audio_capture.py       # WASAPI capture (system + per-app)
      process_audio_capture.py  # Win11 per-process capture
      recorder.py            # Recording state machine
    transcription/
      transcriber.py         # Faster Whisper integration
      diarizer.py            # Speaker diarization (pyannote)
    ai/                      # AI provider integrations (6 providers)
    ui/                      # All UI components
    utils/                   # Config, device enumeration, helpers
  tests/                     # Unit tests
  recordings/                # Output directory

Running Tests

python -m pytest tests/ -v

Tech Stack

Use Cases

• Meeting minutes — Record your Teams or Zoom meetings and get searchable, timestamped transcripts
• Interview recording — Capture job interviews or user research calls with speaker labels
• Lecture capture — Record online lectures or webinars for later review
• Podcast recording — Record remote podcast guests from Discord or Zoom with per-speaker transcripts
• Compliance & documentation — Keep records of client calls with exportable transcripts
• Accessibility — Generate subtitles (SRT) from any recorded call for hearing-impaired participants

Known Limitations

• Windows only — uses WASAPI and Windows COM APIs
• Per-app capture requires Windows 11 Build 22000+
• Per-process COM capture is in active development — the pipeline structure is in place, with packet reading being completed through real-device testing

License

MIT — see LICENSE.