GitHub - varunr89/deskmic

Always-on Windows 11 audio recorder that captures microphone and Microsoft Teams process audio using WASAPI and Application Loopback Capture. Uses Silero VAD with a ring buffer to only save speech segments as WAV files. Includes an async batch transcription pipeline with pluggable backends (local whisper-rs and Azure OpenAI Whisper API), plus LLM-powered daily/weekly email summaries of your transcripts. Lightweight, open-source, single portable .exe.

When recording, deskmic runs a system tray icon with controls to pause/resume recording, open the recordings folder, and quit. It automatically recovers from audio device changes and sleep/wake cycles with exponential backoff.

Quick start

Download the latest deskmic.exe binary and run it:

By default, deskmic records microphone + Teams audio to %LOCALAPPDATA%\deskmic\recordings\, organized by date. A default config is used if no config file exists.

To run on startup:

Interactive setup wizard

For first-time setup, run:

The wizard walks you through four steps:

1. Download Whisper model — choose and download a GGML model (tiny.en, base.en, or small.en) from Hugging Face.
2. Generate config file — create a deskmic.toml next to the executable with sensible defaults.
3. Email summaries (optional) — enter your Azure OpenAI and Azure Communication Services credentials to enable daily/weekly email summaries.
4. Windows startup (optional) — add deskmic to the Windows Startup folder.

If you enable email summaries, the wizard also creates Windows Scheduled Tasks for automatic daily (7 AM) and weekly (Monday 7 AM) summary delivery.

Configuration

deskmic looks for a config file in this order:

1. Path passed via --config <path>
2. deskmic.toml next to the executable
3. %APPDATA%\deskmic\config.toml
4. Built-in defaults

Example deskmic.toml with all options and defaults:

[capture]
sample_rate = 16000
bit_depth = 16
channels = 1

[vad]
speech_threshold = 0.5
pre_speech_buffer_secs = 5.0
silence_threshold_secs = 3.0

[output]
directory = "C:\\Users\\YourName\\AppData\\Local\\deskmic\\recordings"
max_file_duration_mins = 30
organize_by_date = true

[targets]
processes = ["ms-teams.exe"]
mic_enabled = true

[storage]
retention_days = 30
cleanup_interval_hours = 6
# max_disk_usage_gb = 50.0  # optional, no limit by default

[transcription]
backend = "local"       # "local" or "azure"
model = "base.en"       # whisper model name or path

[transcription.azure]
endpoint = ""
api_key = ""
deployment = ""

[transcription.idle_watch]
cpu_threshold_percent = 20.0
idle_check_interval_secs = 30

[summarization]
# deployment = "gpt-4o"                  # Azure OpenAI chat deployment (reuses [transcription.azure] endpoint/key)
# acs_endpoint = "https://your-acs.unitedstates.communication.azure.com"
# acs_api_key = ""                        # or set DESKMIC_ACS_KEY env var
# sender_address = "DoNotReply@your-domain.azurecomm.net"
# recipient_address = "you@example.com"
# system_prompt = ""                      # custom LLM prompt; use {date_label} placeholder

CLI reference

deskmic [OPTIONS] [COMMAND]

Global options:

Commands:

Running deskmic with no subcommand is equivalent to deskmic record.

Transcription setup

Local (whisper-rs)

1. Download a Whisper GGML model (e.g. ggml-base.en.bin) from Hugging Face.
2. Set the model in your config:

[transcription]
backend = "local"
model = "C:\\path\\to\\ggml-base.en.bin"

3. Run deskmic transcribe or deskmic transcribe --watch.

Azure OpenAI Whisper

1. Set up an Azure OpenAI resource with a Whisper deployment.
2. Configure your credentials:

[transcription]
backend = "azure"

[transcription.azure]
endpoint = "https://your-resource.openai.azure.com"
api_key = "your-api-key"
deployment = "whisper-1"

The API key can also be set via the DESKMIC_AZURE_KEY environment variable instead of putting it in the config file.

Summarization setup

The summarize command uses Azure OpenAI to generate an LLM-powered summary of your transcripts and (optionally) emails it via Azure Communication Services (ACS).

Prerequisites

• An Azure OpenAI resource with a chat completion deployment (e.g. gpt-4o). The summarizer reuses the same endpoint and API key from [transcription.azure].
• An Azure Communication Services resource with an Email-verified domain (for email delivery).

Configuration

Add a [summarization] section to your deskmic.toml:

[transcription.azure]
endpoint = "https://your-resource.openai.azure.com"
api_key = "your-azure-openai-key"
deployment = "whisper-1"

[summarization]
deployment = "gpt-4o"
acs_endpoint = "https://your-acs.unitedstates.communication.azure.com"
acs_api_key = "your-acs-access-key"
sender_address = "DoNotReply@your-domain.azurecomm.net"
recipient_address = "you@example.com"

The ACS API key can also be set via the DESKMIC_ACS_KEY environment variable.

Usage

deskmic summarize              # summarize yesterday's transcripts (default: "daily")
deskmic summarize weekly       # summarize the last 7 days
deskmic summarize 2026-02-15   # summarize a specific date
deskmic summarize 2026-02-10..2026-02-14  # summarize a date range (max 90 days)

Summaries are always saved locally as Markdown files under recordings/summaries/, even if email delivery is not configured or fails.

Tip: Run deskmic setup to configure summarization credentials interactively — no manual config editing needed.

Building from source

Requires:

• Rust toolchain (stable)
• Windows 11 SDK (for WASAPI and Application Loopback Capture APIs)

The binary is at target/release/deskmic.exe.

Note: The project compiles on Linux/WSL2 for development (audio capture is stubbed out), but full functionality requires Windows 11.

Legal notice

Recording consent disclaimer: Users are solely responsible for complying with all applicable local, state, and federal laws regarding the recording of audio conversations. Many jurisdictions have two-party (or all-party) consent laws that require all participants to consent before a conversation may be recorded. Use of this software to record conversations without proper consent may be illegal. The authors of deskmic accept no liability for misuse.

License

MIT