Voice to Text

Record audio from your microphone and transcribe it to text using whisper.cpp. The program records until you press Enter, then runs the recording through Whisper and prints the transcription.

Platform: macOS (capture uses AVFoundation). Linux/Windows would require changing the ffmpeg input in code.

Prerequisites

• Go 1.25 or later
• ffmpeg (install via Homebrew: brew install ffmpeg)
• whisper.cpp (included as a submodule; must be built and the model downloaded)

Setup

1. Clone with submodules

   git clone --recurse-submodules <repo-url>
   cd voice-to-text

   If the repo is already cloned:

   git submodule update --init --recursive

2. Build whisper.cpp

   From the project root:

   cd third_party/whisper
   mkdir -p build && cd build
   cmake ..
   make
   cd ../../..

   Ensure the CLI binary is at third_party/whisper/build/bin/whisper-cli. If your build puts the binary elsewhere (e.g. third_party/whisper/main), copy or symlink it to that path.

3. Download a Whisper model

   The app expects the large-v3-turbo model by default at:

   third_party/whisper/models/ggml-large-v3-turbo.bin

   Download it using the whisper.cpp script (from project root):

   mkdir -p third_party/whisper/models
   third_party/whisper/models/download-ggml-model.sh large-v3-turbo third_party/whisper/models

   Other models: You can use a different model by setting WHISPER_MODEL, e.g. for the smaller English base model:

   third_party/whisper/models/ggml-base.en.bin

   Model files are large (large-v3-turbo is ~1.5 GB); they are gitignored.

   Lighter option: For a smaller/faster model, use e.g. ggml-base.en.bin or ggml-small.en.bin and set WHISPER_MODEL accordingly.

Usage

Run from the project root (paths to the CLI and model are relative to it):

go run ./cmd/vtt

Or build and run the binary:

go build -o vtt ./cmd/vtt
./vtt

1. The program starts and prints "Recording... Press ENTER to stop."
2. Speak; when done, press Enter to stop recording.
3. The audio is saved to cmd/vtt/input.wav and then transcribed.
4. The transcription is printed under "--- TRANSCRIPTION ---".

Recordings and generated .wav files are gitignored.

Project layout

• cmd/vtt/main.go – entrypoint: recording via ffmpeg, then transcription via whisper-cli
• third_party/whisper – whisper.cpp as a git submodule (build output and models are local only)

How it works

• Recording: ffmpeg captures from the default system microphone (:0) with AVFoundation, applies light noise reduction, and outputs 16 kHz mono for Whisper.
• Transcription: The saved WAV is passed to whisper-cli with -sns (suppress non-speech tokens) and -l en for cleaner English output.