speak-to

Voice input for CLI tools via fully local Voxtral Mini 4B transcription. Everything runs locally on your machine.

Usage

One-shot mode

Record a voice prompt, transcribe it, and launch a CLI tool with the result:

# Transcribe and pass to claude as the initial prompt
speak-to claude

# Same, with extra arguments forwarded to the client
speak-to claude -- --model sonnet

# Transcribe to stdout (for piping)
speak-to

The flow: load model, record until you press Enter, transcribe, exec the client with the transcribed text as the first argument.

Interactive mode

Wrap any interactive CLI tool in a PTY and trigger voice input at any time with Ctrl+\:

speak-to -i claude

speak-to -i kiro-cli

speak-to -i claude -- --model sonnet

The client runs normally in your terminal. Press Ctrl+\ to start recording, speak, then press Enter. The transcribed text is injected into the client's input as a paste. Press Ctrl+C during recording to cancel.

The model stays loaded in memory between recordings so subsequent transcriptions are fast.

Install

Requires a working Rust toolchain.

cargo install --path .

On first run, the Q4 model (~4GB) is downloaded from HuggingFace and cached locally.

Pre-download the model

speak-to --download

macOS audio permissions

macOS will prompt for microphone access on first use. Grant it to your terminal emulator (iTerm2, Terminal.app, etc.).

Model options

The default is Voxtral Mini Q4 (quantized, ~4GB download, ~700MB memory). For lower word error rate at the cost of ~9GB memory:

speak-to --f32 claude

speak-to --download --f32

You can also point to a local model:

# Q4 GGUF file
speak-to --model-path /path/to/voxtral-q4.gguf claude

# F32 SafeTensors directory
speak-to --model-path /path/to/model-dir/ claude

How it works

One-shot mode records from your mic, transcribes locally on the GPU via burn, and execs the target CLI with the result.

Interactive mode wraps the client in a PTY so speak-to sits between you and the client. It intercepts the trigger keystroke, records and transcribes, then pastes the text into the client's prompt. The client has no idea anything special happened -- it just sees text appear as if you typed it.

Requirements

• macOS (uses CoreAudio for recording and system sounds)
• GPU with wgpu support (Metal on macOS)
• ~4GB disk for the Q4 model (~9GB for F32)

Shoutouts

Transcription is powered by voxtral-mini-realtime-rs by TrevorS, a Rust implementation of Mistral's Voxtral Mini model using the burn framework.

License

Apache-2.0