LM Studio 1.1.3 adds local voice transcription on Linux

The Bionic 1.1.3 update introduces real-time, local speech-to-text capabilities to Linux, matching existing support for macOS and Windows. By converting microphone input directly into prompt text on the user's machine, the desktop application eliminates the need for external cloud APIs or third-party bridge software. This offline transcription feature currently runs on Apple Silicon and NVIDIA GPUs, while AMD GPU compatibility is still in development.

Alongside voice input, the free update integrates llama.cpp 2.38.0 extension packs and expands support for multi-token prediction speculative decoding to accelerate generation on compatible models. It also introduces inline image rendering for agent-created visuals and smoother remote model management via LM Link. However, the company has not disclosed the specific speech model, memory footprint, latency metrics, or supported languages powering the new transcription engine.

For developers and practitioners, this release simplifies the creation of private, voice-controlled local agents. Because the system processes audio segments in real time, users can dictate prompts or notes directly into the chat interface. However, resource management is a key consideration. While Apple Silicon utilizes unified memory, NVIDIA systems must split VRAM among the speech model, the primary language model, and the context cache, which could cause performance bottlenecks during heavy workloads.

Practitioners should note that while LM Studio continues to host its local OpenAI-compatible server at http://localhost:1234/v1, this release does not yet expose a dedicated local audio endpoint like /audio/transcriptions. Transcription remains strictly an in-app feature for now. Furthermore, while local processing secures the initial audio stream, developers must still audit their entire data pipeline, as generated text can still exit the local environment through networked Model Context Protocol servers, plugins, or remote model integrations.