whisper.ihm

Offline speech-to-text transcription tool for long audio files. Built with whisper.cpp and ten-vad for voice activity detection.

Features

• MP3 input with automatic resampling to 16kHz mono
• VAD-based segmentation — splits audio by silence, transcribes each chunk
• Accurate timestamps per segment
• Runs fully offline, no API keys required
• Metal GPU acceleration on macOS (Apple Silicon)

Install

Homebrew (macOS)

brew install tggo/tap/whisper-ihm

From source

Requires Go 1.23+, CMake, Git.

git clone https://github.com/tggo/whisper.ihm.git && cd whisper.ihm
make setup   # clones deps, builds whisper.cpp, downloads model (~3 GB)
make build   # compiles the binary
./whisper-ihm recording.mp3

Usage

Usage: whisper-ihm [flags] <input.mp3>

Flags:
  -model string    Path to GGML model (default "models/ggml-large-v3.bin")
  -lang string     Language code (default "auto")
  -threads int     Number of threads (default: all CPUs)
  -help            Show help

Output format

[00:00:01.200 -> 00:00:05.800] Hello, how are you today?
[00:00:06.100 -> 00:00:09.400] I'm doing well, thank you.

Install from release

Download a pre-built binary from Releases:

# macOS (Apple Silicon)
curl -L https://github.com/tggo/whisper.ihm/releases/latest/download/whisper-ihm-darwin-arm64.tar.gz | tar xz

# Linux (amd64)
curl -L https://github.com/tggo/whisper.ihm/releases/latest/download/whisper-ihm-linux-amd64.tar.gz | tar xz

# Download the whisper model (~3 GB)
mkdir -p models
curl -L -o models/ggml-large-v3.bin \
  https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3.bin

./whisper-ihm recording.mp3

Docker

# Build the image (Linux, CPU-only)
docker build -t whisper-ihm .

# Download model and transcribe
docker run -v $(pwd)/data:/data whisper-ihm -model /data/ggml-large-v3.bin /data/recording.mp3

The Dockerfile uses a multi-stage build: golang:1.23-bookworm for building (clones whisper.cpp + ten-vad, compiles with CGO), debian:bookworm-slim for runtime.

Build details

• -trimpath strips local filesystem paths from the binary
• macOS builds include Metal GPU acceleration
• Linux/Docker builds are CPU-only

How it works

1. Decode MP3 to PCM, resample to 16kHz mono
2. Run VAD (ten-vad) to detect speech segments, split on ~500ms silence gaps
3. Feed each segment to whisper.cpp with timestamp offsets
4. Print [start -> end] text for each whisper segment

License

MIT