Local, live English captions for the audio playing through your default PulseAudio/PipeWire output device. It captures that output's monitor source, runs the Parakeet TDT v2 model locally, and writes captions to the terminal.

Everything runs on your machine after the one-time model download. No audio is uploaded anywhere.

Requirements

Linux on amd64 / x86_64
Go 1.26.2 or newer
pactl and parec available on PATH (on Debian/Ubuntu: sudo apt install pulseaudio-utils)
curl and tar to download the models
A working PulseAudio-compatible server, including PipeWire's PulseAudio compatibility layer

The application listens to the monitor of the current default audio sink. Check it with:

pactl get-default-sink

Setup

Clone the project, then download the required models once:

git clone <your-repository-url>
cd audio-listen
./scripts/download-models.sh

The download installs the Parakeet TDT v2 int8 model and Silero VAD into models/. They are intentionally not committed to Git because they are large.

Run

Use VAD mode for normal live captions. It shows a temporary … caption while speech is active, then prints a timestamped final caption after 0.5 seconds of silence:

go run . --mode=vad

Use fixed mode to decode exact time slices, even through silence:

go run . --mode=fixed --chunk-duration=1s

Build a reusable executable instead of using go run:

go build -o audio-listen .
./audio-listen --mode=vad

Press Ctrl+C to stop. Pending audio is flushed before exit.

Options

--mode is required. All other options have defaults.

Option	Default	Description
`--mode=vad`	—	Speech-aware captions. This is the recommended everyday mode.
`--mode=fixed`	—	Decode every fixed-duration audio chunk.
`--chunk-duration=1s`	`1s`	Fixed chunk length, or VAD preview refresh interval. Go duration syntax is used, e.g. `500ms`, `1s`, `2s`.
`--models-dir=models`	`models`	Directory containing `parakeet-tdt-v2/` and `silero_vad_v5.onnx`.
`--threads=2`	`2`	CPU threads used by Parakeet and VAD. Increase only if your CPU has spare capacity.
`--preview-threshold-dbfs=-45`	`-45`	RMS level that begins a provisional VAD caption. Raise it (for example, `-35`) if background audio starts previews; lower it (for example, `-55`) if quiet speech does not.

Examples:

# Faster fixed chunks
go run . --mode=fixed --chunk-duration=500ms

# Fewer VAD preview redraws and four CPU threads
go run . --mode=vad --chunk-duration=2s --threads=4

# Store or reuse models outside this repository
go run . --mode=vad --models-dir=/path/to/models

There are no environment variables to configure. All supported configuration is supplied with command-line options.

Output and troubleshooting

Final captions go to standard output in this form:

[12.34s-15.67s] This is a completed caption.

Startup messages and overload warnings go to standard error, so you can save final captions separately:

go run . --mode=vad > captions.txt

Missing model files: run ./scripts/download-models.sh again.
pactl or parec not found: install pulseaudio-utils and start a PulseAudio-compatible audio server.
No captions while audio plays: verify the default sink with pactl get-default-sink; the app records that sink's .monitor source.
Overload warning: recognition is behind real time. Try --threads=4 (or another suitable value) or --chunk-duration=2s.

Verify changes

go test ./...
go build ./...