Speakrail

Speakrail is a full-duplex voice assistant that runs locally on a single NVIDIA RTX graphics card.

Speakrail is a full-duplex voice assistant that listens while it talks, allowing interruptions, backchannels, and pauses mid-sentence without being cut off. It uses a turn-taking head on Voxtral's hidden states and Gemma 4 12B with a LoRA for decision-making. The system runs locally using Docker on Linux with an NVIDIA RTX 3090, 4090, or similar card. Built-in tools include weather, timers, notes, lists, calculator, unit conversion, and optional web search. The package requires about 75 GB of disk space and supports web browser access via localhost.

Speakrail at a glance

Best for: Software engineer
Runs on: Web, Linux, self-hosted
Pricing: Free
Released: October 2026
Category: AI audio tool
Website: github.com/speakrail/speakrail

Key features

Full-duplex conversation with interruption handling
Turn-taking head for speech state detection
Gemma 4 12B model for decision-making
Built-in tools for weather, timers, and notes
Peek and speculation for early reply generation
Listening notes for long or corrected requests

Use it for

Count workout reps while talking to the assistant
Manage timers and weather updates through voice
Execute local voice assistant tasks without cloud services

Alternatives to Speakrail

Qwen Audio Agent — Separates voice conversation from task execution for continuous interaction.
walkai-talkai — Control AI coding agents remotely using any handheld walkie-talkie.
Speech-to-Speech — Fully modular, swappable components with local or hosted execution options.
AssemblyAI — Industry-leading speech models deliver high accuracy and low latency for voice applications.
Pipecat — Builds real-time voice and multimodal agents with frame-level control.
Willow — Fast, private, open source voice control using local hardware.
talkflow — Transcribes speech entirely on your device with no word limits or cloud accounts
Auricle — Runs entirely in the browser via Rust and WebAssembly with low latency and high accuracy.
FunASR — It supports offline, streaming, and edge deployment with multiple models.
whisper.cpp — It provides high-performance, offline speech recognition on diverse hardware.
Babytalk — Runs entirely on-device with no cloud or internet required.
sherpa-onnx — Runs complex audio AI tasks offline on local devices and embedded hardware.

All Speakrail alternatives